feat(gemma4 ARF): infrastructure - wrapper + Gemma3MLP skip_all_reduce + auto-enable (PR-A/2)

[pyc96]

Summary

PR-A of a 2-PR stack that adds FlashInfer AllReduce + RMSNorm fusion infrastructure for Gemma-4. Does not change any runtime behavior on its own — the wrapper is added but not wired. PR-B (#TBD) wires it into the post-attention site in `Gemma4DecoderLayer.forward`.

Background

vLLM's `fuse_allreduce_rms` Inductor pass is technically enabled for Gemma-4 at compile mode O2 (per vllm/config/vllm.py:122-142) but never matches Gemma-4's residual flow because Gemma uses `RMSNorm(x) + residual` rather than the Llama-style two-arg `RMSNorm(x, residual)` the pattern looks for. SGLang already has the `flashinfer_allreduce_residual_rmsnorm` direct-call API (used by Qwen3-MoE, DeepSeek-V3, GLM4-MoE etc.). Adding it to Gemma-4 gives SGLang a fusion vLLM doesn't actually deliver on Gemma-4 today.

What's in PR-A

File	Change
`python/sglang/srt/layers/gemma4_fused_ops.py`	New function `gemma4_arf_rmsnorm_residual_scalar(x, weight, residual, scalar, eps, use_attn_tp_group=True)` that wraps `flashinfer_allreduce_residual_rmsnorm` and applies a `* scalar` tail. Falls back to `tensor_model_parallel_all_reduce + gemma_rmsnorm_residual_scalar` on any predicate-False / non-CUDA / non-2D / FlashInfer-unavailable signal.
`python/sglang/srt/models/gemma3_causal.py`	`Gemma3MLP.forward` (= `Gemma4MLP.forward` via alias) now accepts `skip_all_reduce: bool = False`, threaded through to `down_proj`. Default `False` preserves current behavior for every other caller.
`python/sglang/srt/server_args.py`	Adds `Gemma4ForCausalLM` + `Gemma4ForConditionalGeneration` to the `flashinfer_allreduce_fusion` auto-enable allow-list. Same preconditions as the existing 13 archs (SM90/100, TP>1, single-node, not H20, no DP-attn, no MoE-A2A).
`test/registered/unit/layers/test_gemma4_arf_ops.py`	4 unit tests with FlashInfer + all-reduce fully mocked (runs on CPU): success path, fallback path, predicate-off path, non-cuda fallback. All pass.

What this PR does NOT change

• `gemma4_causal.py` is unchanged — the new wrapper is not called yet.
• No runtime behavior change for any deployment that doesn't pass `--enable-flashinfer-allreduce-fusion` AND wait for PR-B to wire the call site.
• The auto-enable list change does mean a Gemma-4 server on TP=2 H100 will log `"Auto-enabling FlashInfer AllReduce Fusion on SM90/SM10X for Gemma4ForCausalLM"` at startup, but without PR-B's wiring nothing actually fires.

Tests

```
$ python test/registered/unit/layers/test_gemma4_arf_ops.py
....
Ran 4 tests in 1.053s
OK
```

Test	What it verifies
`test_success_path_uses_flashinfer_and_applies_scalar`	wrapper returns `norm_out * scalar` from mocked FlashInfer; AR helper / fallback kernel NOT invoked
`test_fallback_when_flashinfer_returns_none`	wrapper falls back to `AR + kernel` sequence when FlashInfer returns `(None, None)`
`test_predicate_off_uses_fallback_directly`	wrapper takes fallback without calling FlashInfer when predicate is False
`test_non_cuda_input_takes_fallback`	CPU tensors short-circuit through fallback (is_cuda gate)

Diff noise note

`server_args.py` shows 327 lines changed but only 9 are mine — the rest is auto-format reflow of unrelated `assert` statements. Real change: 4 new lines in the auto-enable list (`"Gemma4ForCausalLM"`, `"Gemma4ForConditionalGeneration"`) + 5-line comment.

Stack

• PR-A (this PR): infrastructure
• PR-B: wire `gemma4_arf_rmsnorm_only` into `Gemma4DecoderLayer.forward` at the post-attention site (Site perf(gemma4 MTP): single-launch fused router (topk + softmax + scale) #1) — the only site in Gemma-4's residual flow that mathematically matches FlashInfer's `kARResidualRMSNorm`

Stack base: `pyc/sota-gemma4-31b-mm-disabled` @ `3a3195b30`

Plan: `.humanize/yoco-gemma4/refined-plan.md` (the ARF stack inherits the same review-plan structure)

---

CI States

Latest PR Test (Base): ❌ Missing run-ci label -- add it to run CI tests.
Latest PR Test (Extra): ❌ Blocked -- run-ci is required first.