perf(musa): optimize Qwen2 decode for CosyVoice3

[yeahdongcn]

Summary

• tune the MP31 small-hidden add-RMSNorm launch used by the Qwen2 talker;
• fuse the exact Qwen2 BF16 RoPE + NHD KV-cache update chain into one MUSA
  Triton launch;
• apply the graph rewrite atomically before vLLM splitting so layer-name
  strings remain outside Inductor graphs and compiled pieces stay reusable;
• keep the upstream KV split and all unsupported model, graph, dtype, prefill,
  and tensor-layout paths on the existing implementation.

This PR changes vLLM-MUSA only. vLLM-Omni is an integration test path, not an
implementation dependency or patch target.

Why the rewrite is pre-split

The first post-grad prototype fused correctly but caused about 95.6 seconds of
cold compilation because 24 layer-name strings produced non-reusable pieces.
The accepted raw-FX rewrite requires all 24 exact sites to match before any
mutation, adds the fused outer op as a split, and retains the original KV split
as a fail-closed fallback.

The accepted path restores compile plus initial warmup to 5.90 seconds on the
development container and produces only four reusable compile artifacts. It is
exact-gated and default-on; no serving flag is required. The environment switch
is retained only as an A/B and emergency-disable escape hatch.

Performance

Single S5000, BF16, FlashAttention-3, normal vLLM compile/FULL graph capture,
357 input tokens, fixed 1024 output tokens (40.96-second audio proxy),
temperature 1.0, repetition penalty 1.0001. Two alternating same-host rounds,
three measured requests per cell after one warmup:

Variant	Runs	Mean RTF	Median RTF
RoPE+KV fusion	6	0.141548	0.141559
baseline on the same branch	6	0.145836	0.145760

Combined mean gain: 2.94%; combined median gain: 2.88%. Both paired rounds are
positive. A deterministic token-identical A/B improves mean RTF from 0.077270
to 0.067954 (12.06%), isolating launch-heavy LLM decode from sampling cost.

A candidate MUPTI activity trace confirms the production path retains all
three native BF16 TCE GEMM families with the expected 2:1:1 call ratio. The
accepted fused kernel averages 4.445 us over the captured startup, graph,
warmup, and measured-request scope. Generated Inductor artifacts use external
addmm/mm kernels and do not Triton-lower the GEMMs.

Validation

• focused FX/backend/config/real-S5000 kernel suite: 19 passed;
• default-on integration: matched 24/24 layers, exercised a 357-token prefill
  fallback, and completed 256 deterministic decode tokens with checksum
  953984;
• deterministic 1024-token candidate and baseline: identical checksum 15219200
  in all six measured requests;
• broader Python slice: 117 passed, 2 skipped, 3 failed; the same three
  unrelated FP8-MoE/format-patch failures reproduce from the pre-overlay source
  archive;
• formatting and lint pass for all new/changed files except three unchanged
  baseline findings in flash_attn.py.

Not run / Draft status

• full synthesized-waveform listening and quality validation belongs to the
  downstream CosyVoice integration provider;
• project CI and maintainer review are still pending.

Keep this PR in Draft until those external gates complete.