[Bench][SM70] Repair #377 without slowing GLM NVFP4

[yangzhuxinyzx]

Purpose

Repair #377 from the latest audited main while retaining its useful GLM-5.3
and NVFP4 benchmark coverage without forcing production onto a slower
accumulation tree.

Audit decision

• Keep the default-on measured NVFP4 selector for the exact GLM W13 shape.
• Remove the shape-specific default-split override. On one V100 with identical
  random input/weights:
  • default split-8 median: 51.200 us;
  • measured split-12 median: 48.128 us (6.0% faster);
  • changed FP16 elements: 37/8192;
  • maximum absolute difference: 0.125;
  • relative L2: 1.642e-5;
  • cosine: 0.99999917.
• Treat that accumulation-order delta as numerical diagnostics rather than
  task-quality evidence; retain a diagnostic tuning opt-out through the
  existing environment control.
• Keep random-data, grouped-MoE, output-hash, and tensor-dump support in the
  existing NVFP4 microbenchmark.
• Keep the new GLM FP16 Dense/Indexer candidate screen benchmark-only; it does
  not alter production dispatch.
• Make the base and eviction-policy Triton candidates genuinely distinct,
  lazy-load the compiled SM70 wrapper so --help works in a clean checkout,
  and repair the accelerator/static-format gates.

Validation

• Real V100 grouped-MoE shape ran successfully and reproduced seed-17 hash
  47836839b542fb73494caa64adc14cd660e38c535c4a5e67e16d3a763196dac7.
• Corrected FP16 benchmark candidates ran on V100; the N32/K4096 smoke reported
  base/evict/cg at 5.530/6.656/7.270 us with exact outputs for that input.
• Eight focused GLM ModelOpt NVFP4 admission tests passed.
• Python compile and clean-checkout CLI help passed.
• Changed-file pre-commit passed, including Ruff, formatting, clang-format,
  mypy, forbidden-import, and accelerator API gates.
• The migration control document records the selector decision and exact
  numerical/performance evidence.
• The repaired source is based on public main at
  b410e59eb9a1283bb20c239bac52c6ba88459ea2; [Bench][SM70] Add GLM NVFP4 quality and performance screens #377 head
  cd1bc99f0404dbac88c37272b0284916a07db373 is preserved as ancestry without
  changing the audited tree.
• All repair commits are DCO signed.

[yangzhuxinyzx]

审计完成并锁定 bb68898c09c33f428c19a6616ee9a5d4a342aef0（tree 3395bf1b98f9d521e56b19d2488dfca3e80a2bbc）。

• 远端完整 pre-commit 通过。
• 8 个聚焦 GLM ModelOpt NVFP4 测试通过。
• V100 实测动态选择器中位延迟 48.128 us，相比默认 split-8 的 51.200 us 快约 6.0%。
• 数值差异为 37/8192 个 FP16 元素，max abs 0.125、relative L2 1.642256e-5、cosine 0.9999991655；重复动态输出哈希稳定。
• 修复后的 PR 不再修改生产 C++ 路由，只增加可复现的性能/质量筛查与迁移记录。

按项目策略保留默认开启的性能路径，准备合并回 #377。