perf: fuse SiLU and FP8 group quant before DeepGEMM

[yeahdongcn]

Summary

This is the first PR in a four-PR MUSA kernel stack:

1. MUSA-0759 (this PR): fuse SiLU and row-major group-128 FP8 activation quantization before MATE DeepGEMM.
2. MUSA-0760: fuse residual-add RMSNorm and row-major group-128 FP8 activation quantization.
3. MUSA-0763: add guarded chunked min-p sampling.
4. MUSA-0765: add guarded uniform top-k=50 renormalization.

Review/merge order: #109 -> #110 -> #111 -> #112.

The compiler pass is selected automatically for the validated Qwen3-8B-FP8,
single-S5000 contract. Unsupported model, parallelism, shape, dtype, device,
or quantization contracts keep the existing path. No campaign- or
fusion-specific environment control is introduced or read to select this
path; guarded automatic dispatch is the selection mechanism.

Exact source:

• base: v0.24.0-dev at 68b5ef7a7eac9959dd71f600f7df2ebd756de1f8
• head: 7d6dd02c4b6a45e00d0131fd42b885030d7891c2
• head ref: xd/musa-0759-silu-deepgemm
• pinned vLLM: ee0da84ab9e04ac7610e28580af62c365e898389

Validation

Historical per-ticket performance evidence

The figures below are retained from the earlier MUSA-0759-only serving
campaign. They predate the current environment-independent and logical-device
follow-ups, so they are directional historical evidence, not a result from the
current exact head and not the final combined four-PR A/B. The run used
Qwen3-8B-FP8, TP1, S5000, normal Inductor,
FULL_AND_PIECEWISE graph capture, greedy decoding, no MTP/speculative
decoding, and separate candidate/native-control source trees. openai-chat
adds eight Qwen template tokens, so --random-input-len 4096 measures 4104
input tokens; every request generated exactly 1024 output tokens.

concurrency	mean TTFT change	mean TPOT reduction
1	+1.89%	+0.38%
4	-4.49%	+0.80%
16	+2.39%	+0.77%
64	+0.84%	+0.32%

TPOT improved in all four historical cells. TTFT was noisy, especially at
concurrency 4, and is not presented as a general win.

Superseded descendant-stack evidence

The former descendant head f53cf6cbbd7f4d85e82270a86db1938c713d948a
passed dense, MoE, and MLA functional serving checks. That SHA is no longer the
final stack, and the same host-81 campaign had a foreign ExecBench workload,
so its combined latency rows are rejected and none of this evidence is used as
an exact-head readiness claim.

Exact-final MUSA-0767 gate

Status: EXACT-HEAD FUNCTIONAL REGRESSION COMPLETE; PERFORMANCE GATES
PENDING -- keep this PR in draft.

• final stacked head: 95c3661c8658a861a1891f981c81dcadabea0654
• final source tree: a766855ed712d61f81acb6b11d456ea5a128cac5
• source archive SHA-256:
  03dbe0526f72d4dcba6b743c500d4d5eaaf9676e263f9bd91cbfd265db140feb
• source manifest SHA-256:
  3ab55a222ea7fc8324c611bda9d3d842bc94850bbc7ccf4d869d0692132768a7
• exact environment and build: 8x MTT S5000, driver 3.3.5-server,
  MUSA runtime 50200, torch==2.9.1.post1,
  torch_musa==2.9.1.post1+4c646b6, torchada==0.1.75,
  MATE/FlashAttention3 0.2.4, candidate extension SHA-256
  d0baaa8ba11981438a549115d50e89c10ec566dfac1ba27c0ccdc8364c5c2bd7
• focused native correctness and graph-capture tests: 60 passed in 44.72s
• non-identity logical-device sampling smoke: PASS with
  MUSA_VISIBLE_DEVICES=4,5; physical devices 4 and 5 were exposed as
  logical devices 0 and 1, and the smoke selected both optimized samplers on
  logical device 0
• compiled/captured Qwen dense, Qwen MoE, and DeepSeek MLA regression:
  PASS, 3/3 rows and 10/10 semantic requests (Qwen3-8B-FP8 TP1
  FLASH_ATTN, Qwen3.5-35B-A3B-FP8 TP4 FLASH_ATTN, and
  DeepSeek-V2-Lite-Chat-FP8 TP1 FLASHMLA; Inductor,
  FULL_AND_PIECEWISE, MTP off)
• clean-node combined Qwen3-8B-FP8 candidate/control result at exact
  4104/1024 and concurrency 1/4/16/64, without MTP:
  PENDING -- not present in this regression-only bundle
  • the first control-only attempt is rejected because default prefix caching
    plus repeated deterministic prompts invalidated cold-prefill TTFT; every
    A/B leg will restart with prefix caching disabled
• Qwen3.5-2B and Qwen3.6-27B exact-token widening A/B evidence:
  PENDING -- not present in this regression-only bundle
• regression evidence bundle SHA-256:
  ea1b38559be0458343b85f3cd6bba275d320d2b1c729071f455e2dccb04c5272;
  its focused monitor verdict was FAIL because foreign ExecBench GPU work
  was present, so the clean performance-monitor gate remains PENDING

Remaining gate

The exact-head functional regression is complete. This PR remains draft until
the combined and widening A/B plus clean performance-monitor gates are
complete and accepted. The stacked PRs must be reviewed and merged in order.

[yeahdongcn]

Superseded by #115, which consolidates the four logical kernel changes on the current v0.24.0-dev head (4416dc24) while preserving per-change commits. This draft and its historical evidence remain available for traceability; it is not a current-head readiness claim.