perf(moe): add Kimi-K3 B1 prepared A8W4 pipeline

[JohnQinAMD]

perf(moe): add Kimi-K3 B1 prepared A8W4 pipeline

Summary

Add a narrow gfx950 Kimi-K3 batch-one MoE pipeline that owns route metadata,
MXFP8 activation quantization, tuned A8W4 expert GEMMs, SiTU, and weighted BF16
reduction behind one typed request.

The fast path is fail-closed. It requires the exact native TP8 checkpoint
layout, validated kernel registry entry, routing semantics, device, dtypes, and
runtime features. Unsupported inputs return to the existing path.

Design

• Generate stable-priority top-16 route metadata for 896 experts while
  preserving exact incumbent tie behavior.
• Overlap the one-row MXFP8 preparation with route selection.
• Register the independently validated 14-wave stage-1 schedule as one explicit
  production variant rather than widening the generic tuning combinations.
• Reuse the tuned stage-1 A8W4 SiTU/quant pipeline and explicitly select the
  native-384 stage-2 atomic weighted reduction.
• Keep eligibility in one named public predicate, split internally into native
  tensor layout, model configuration, and runtime-feature contracts.
• Expose immutable request and metadata dataclasses so the consumer owns
  prepared routing exactly once.
• Leave generic registry selection and non-Kimi paths unchanged.

Performance

MI355X, one GPU microbenchmark, rotating through all 896 experts:

Seed	Candidate	Native incumbent	Speedup
111	18.8658 us/layer	27.5647 us/layer	1.4611x
223	18.8490 us/layer	27.5779 us/layer	1.4631x

On the full TP8 Kimi-K3 endpoint (B1, 8,192 input tokens, 1,024 output tokens,
no speculative decoding), two measured runs were 69.7324 and 69.8257 tok/s.
The 69.7791 tok/s mean is 8.56% above the hash-pinned 64.2744 tok/s parent and
reduces mean TPOT by 1.2275 ms.

The final clean signed commit was rebuilt as a separately hash-pinned image
and confirmed at 69.4669 and 69.8105 tok/s (69.6387 mean), within 0.21% of
the accuracy-tested image.

Correctness and tests

• Native-layout kernel correctness: 20/20 cases, 1,000 graph replays.
• Minimum cosine similarity: 0.9999848008.
• Maximum relative L2: 0.0055199412.
• NaN/Inf: zero.
• Focused AITER contract and GPU tests: 31 passed.
• Ruff format/check and git diff --check: passed.
• Frozen GSM8K completion contract: 1,256/1,319 (0.952237), one invalid
  response. The hash-pinned parent scored 1,244/1,319 (0.943139), with two
  invalid responses; paired outcomes were 32 candidate wins and 20 losses.

Scope and dependencies

This is stacked on JohnQinAMD/aiter-amd:perf/kimi-k3-preroute-bf16.
It intentionally excludes the rejected direct-BF16 stage-1 experiment and does
not change the generic Kimi-K3 tuning CSV.

The matching vLLM integration is a separate PR against
JohnQinAMD/vllm-amd.

AI assistance

OpenAI Codex assisted with implementation, profiling analysis, test
construction, and PR preparation. The human submitter reviewed the changed
lines and is responsible for the design, measurements, and submission.

[JohnQinAMD]

Superseded by DCO-clean, patch-equivalent replacement #22. The replacement adds the required Signed-off-by trailers without force-pushing this reviewed branch; performance, correctness, and accuracy evidence is preserved.