Give every subcommand the same runtime flags

[justinchuby]

The problem

serve answers the same four questions as generate, run and transcribe --
which backend, which device, how much VRAM, how much host RAM -- in a different
vocabulary. Before this change:

flag	generate	run	transcribe	serve
--backend	yes	yes	yes	no
--device	yes	yes	yes	no (spelled --native-device)
--host-ram-limit	yes	yes	yes	no
--cpu-cores	yes	yes	yes	no
model positionally	yes	yes	n/a	no (--model only)

So a command line learned from one subcommand is rejected by another.
onnx-genai serve MODEL --backend native --device cuda:6 -- the obvious
extrapolation from run -- fails twice in a row, once on the positional model
and once on --backend, and clap's usage text lists neither. The flag that
does work, --native-device, appears in no other subcommand's help, so there
is nowhere to discover it from.

The change

The flags now live once, in onnx-genai-server::runtime_args, and are
flattened into all four subcommands. That crate is the lowest layer both the
unified CLI and the standalone server binary already depend on, and it already
carries clap with the env feature. A flag added there appears everywhere with
the same name, grammar and help text, so the surfaces cannot drift apart again.

Duplicated definitions removed along the way: EngineArgs, CpuArgs, the
device enum, and the --backend, --device and memory-limit parsers each
existed twice, and parse_native_device had two implementations that disagreed
about whether auto and gpu were valid.

Behaviour changes

Each has a test:

• serve accepts the model positionally. Both spellings share the model-source
  arg group, so giving it twice is refused while parsing rather than after the
  process has started.
• Naming a --device implies the native backend on every subcommand. serve
  already worked this way through --native-device; previously
  generate --device cuda:0 parsed, ran on ORT and ignored the device.
• --vram-limit and friends gain the env fallbacks serve already had.
• --model-id alongside --models-dir/--models-config is an error instead of
  being silently ignored, which hid typo'd invocations.

--native-device keeps working and is hidden from help, so existing service
units are unaffected. --device wins if both are given.

Verification

• cargo test -p onnx-genai-server --lib --features native-backend — 16 tests
  across cli:: and runtime_args::, all passing; full lib 237 passed.
• cargo test -p onnx-genai-cli --lib --features native-cuda — 118 passed.
• cargo fmt --all --check and cargo clippy clean.
• End to end on a real native CUDA model: onnx-genai serve ~/models/... --backend native --device cuda:6 --addr 127.0.0.1:18082
  loads and serves a chat completion, which is the invocation that failed before.

[github-actions]

🔴 Benchmark Regression Detected

Comparison of criterion micro-benchmarks: PR head vs merge-base, measured on the same runner in the same job (base first → PR second).

ℹ️ Absolute times are informational only — they vary with runner load. The % change column is the reliable signal because both sides ran under identical conditions.

Status	Scenario	Base	PR	Change
🔴	block_quantized_matmul_cached_dense/mxfp4_uncached_dequant_each_call/1x1024x1024	519.85 µs	1.35 ms	+158.9%
🔴	matmul/large_generic_bf16_threads=8/32x1024x1024	1.30 ms	2.00 ms	+54.0%
🔴	block_quantized_matmul_cached_dense/mxfp4_cached_dense_repeated_call/1x1024x1024	40.76 µs	59.45 µs	+45.8%
🔴	gather/medium_f16_threads=1-internal/32768	2.67 µs	3.71 µs	+38.8%
🔴	matmul/medium_generic_f16_threads=1/32x512x512	30.08 µs	40.92 µs	+36.0%
🔴	block_quantized_matmul_cached_dense/mxfp4_preexpanded_dense_oncelock_like_proxy/1x1024x1024	40.63 µs	53.29 µs	+31.1%
🔴	block_quantized_moe_cached_dense/mxfp4_cached_dense_expert_repeated_call/rows=1,H=256,I=256,E=4,top_k=1	55.68 µs	72.99 µs	+31.1%
⚠️	matmul/medium_generic_f16_threads=8/32x512x512	29.49 µs	37.80 µs	+28.2%
⚠️	gather/medium_bf16_threads=1-internal/32768	2.57 µs	3.11 µs	+21.1%
⚠️	matmul/large_generic_f32_threads=1/32x1024x1024	9.19 ms	11.03 ms	+20.0%
⚠️	block_quantized_moe_cached_dense/mxfp4_uncached_expert_dequant_each_call/rows=1,H=256,I=256,E=4,top_k=1	354.01 µs	424.52 µs	+19.9%
⚠️	sampling_latency/top_k_per_token	53.51 µs	63.67 µs	+19.0%
✅	matmul/large_generic_f16_threads=1/32x1024x1024	80.35 µs	90.80 µs	+13.0%
✅	gather/small_f16_threads=1-internal/4096	529.5 ns	597.5 ns	+12.8%
✅	matmul/large_generic_f16_threads=8/32x1024x1024	95.66 µs	107.41 µs	+12.3%
✅	sampling_latency/top_p_per_token	390.09 µs	436.73 µs	+12.0%
✅	matmul/large_generic_bf16_threads=1/32x1024x1024	2.01 ms	2.23 ms	+10.7%
✅	matmul/small_generic_f32_threads=8/1x256x256	59.78 µs	65.87 µs	+10.2%
✅	gather/small_bf16_threads=1-internal/4096	478.0 ns	524.9 ns	+9.8%
✅	reduce_mean/large_f32_threads=1-internal/262144	1.07 ms	1.17 ms	+9.7%
✅	qwen3_sampling_processors/top_p_full_sort_after_top_k_baseline	3.62 ms	3.93 ms	+8.5%
✅	qwen3_sampling_processors/top_p_fast_after_top_k	564.64 µs	598.16 µs	+5.9%
✅	gather/small_f32_threads=1-internal/4096	970.7 ns	1.02 µs	+5.3%
✅	matmul/medium_generic_bf16_threads=1/32x512x512	550.64 µs	570.25 µs	+3.6%
✅	gather/medium_f32_threads=1-internal/32768	4.14 µs	4.27 µs	+3.3%
✅	sampling_latency/min_p_per_token	232.69 µs	237.97 µs	+2.3%
✅	matmul/small_generic_bf16_threads=8/1x256x256	39.03 µs	39.76 µs	+1.9%
✅	logit_processing/seven_processor_chain_per_step	322.18 µs	325.54 µs	+1.0%
✅	matmul/large_generic_f32_threads=8/32x1024x1024	4.78 ms	4.80 ms	+0.4%
✅	grammar_masking/llguidance_compute_mask/32	76.89 µs	76.54 µs	-0.5%
✅	matmul/small_generic_bf16_threads=1/1x256x256	39.78 µs	39.13 µs	-1.6%
✅	matmul/medium_generic_f32_threads=1/32x512x512	2.62 ms	2.55 ms	-2.4%
✅	tokenization/decode_tokens_per_second	7.64 ms	7.38 ms	-3.4%
✅	matmul/small_generic_f32_threads=1/1x256x256	48.92 µs	46.72 µs	-4.5%
✅	gather/large_bf16_threads=1-internal/131072	16.74 µs	15.98 µs	-4.5%
✅	kv_cache/alloc_dealloc_pages	40.85 µs	38.92 µs	-4.7%
✅	qwen3_sampling_processors/top_k_full_sort_baseline	2.80 ms	2.66 ms	-5.0%
✅	qwen3_sampling_processors/top_k_top_p_full_sort_baseline	6.56 ms	6.20 ms	-5.6%
✅	sampling_latency/greedy_per_token	3.75 µs	3.49 µs	-7.1%
✅	reduce_mean/medium_f32_threads=1-internal/65536	284.81 µs	261.96 µs	-8.0%
✅	matmul/small_generic_f16_threads=1/1x256x256	48.20 µs	43.79 µs	-9.2%
✅	qwen3_sampling_processors/top_k_partial_selection	189.09 µs	171.75 µs	-9.2%
✅	qwen3_sampling_processors/top_k_top_p_fast	827.79 µs	745.41 µs	-10.0%
✅	tokenization/encode_tokens_per_second	460.34 µs	410.32 µs	-10.9%
🟢	gather/large_f16_threads=1-internal/131072	15.55 µs	13.21 µs	-15.0%
🟢	matmul/medium_generic_bf16_threads=8/32x512x512	579.98 µs	487.72 µs	-15.9%
🟢	gather/large_f32_threads=1-internal/131072	37.34 µs	30.02 µs	-19.6%
🟢	add/small_f16_threads=1-internal/1024	541.7 ns	419.1 ns	-22.6%
🟢	matmul/medium_generic_f32_threads=8/32x512x512	1.32 ms	1.01 ms	-23.5%
🟢	matmul/small_generic_f16_threads=8/1x256x256	45.01 µs	34.18 µs	-24.1%
🟢	add/large_bf16_threads=1-internal/4194304	2.09 ms	1.55 ms	-25.9%
🟢	add/large_f16_threads=1-internal/4194304	2.13 ms	1.54 ms	-27.8%
🟢	reduce_mean/small_f32_threads=1-internal/4096	19.52 µs	14.09 µs	-27.9%
🟢	add/small_bf16_threads=1-internal/1024	644.4 ns	447.4 ns	-30.6%
🟢	add/small_f32_threads=1-internal/1024	329.5 ns	203.6 ns	-38.2%
🟢	add/large_f32_threads=1-internal/4194304	983.92 µs	565.98 µs	-42.5%
🟢	add/medium_f32_threads=1-internal/262144	42.32 µs	22.35 µs	-47.2%
🟢	add/medium_bf16_threads=1-internal/262144	201.23 µs	96.10 µs	-52.2%
🟢	add/medium_f16_threads=1-internal/262144	277.02 µs	99.19 µs	-64.2%

Visual flags: ⚠️ ≥ 15% slower, 🔴 ≥ 30% slower — calibrated against measured runner noise (~27% worst-case on multi-threaded matmul)

Host info

CPU: Apple M1 (Virtual)
Cores: 3
OS: Darwin 25.5.0 arm64
Rust: rustc 1.97.1 (8bab26f4f 2026-07-14)
Load avg: { 3.51 3.60 5.91 }

What this cannot catch

• Regressions in code paths not covered by these benchmarks (e.g., end-to-end decode with a real model)
• Sub-threshold regressions that compound over multiple PRs
• Performance changes that only manifest under GPU execution
• Latency changes in the ORT integration path (these benchmarks exercise the native Rust kernels)