[Core][Kernel][SM70] Fix PP profiling and compact MXFP4 shapes

[yangzhuxinyzx]

Purpose

Supersede the accepted engine-level subset of #283 without carrying its rejected model-workload experiments into main.

• unwrap non-last pipeline-stage dummy outputs through the semantic hidden_states tensor
• admit the existing default-off compact MXFP4 grouped operator for two additional exact SM70 shapes
• keep runtime admission independent of model name, checkpoint, model_type, or architecture identity

The original #283 QPN8 candidate showed repeated baseline-correct-to-candidate-wrong quality regressions. Its static PP candidate recovered only about 0.034 ms/token at the endpoint. Neither is included here.

Test plan and result

• changed-file pre-commit: passed
• CPU: 28 passed, 3 GPU-only skipped
• CUDA 12.8 full extension build, target architecture 7.0 only: passed
• V100 exactness and CUDA Graph replay for K4096/N1024 and K512/N4096: 2 passed, bitwise equal to the per-expert fallback
• final-source V100 microbenchmark:
  • K4096/N1024: 0.151521 ms to 0.040428 ms, 3.748x
  • K512/N4096: 0.065004 ms to 0.022538 ms, 2.884x

No full-model E2E run is claimed or required for this source/operator audit.

[yangzhuxinyzx]

审计结论：可合并。

• 运行准入只基于 SM70、量化格式、张量形状与 PP 中间张量语义，不依赖模型/检查点/架构身份。
• 变更文件 pre-commit 全通过；CPU 28 passed、3 GPU-only skipped。
• CUDA 12.8 SM70 扩展编译通过，产物仅含 sm_70；V100 两个精确 shape 均 bitwise 一致并通过 CUDA Graph replay。
• 当前主线微基准：4096x1024 为 0.151521 -> 0.040428 ms（3.748x）；512x4096 为 0.065004 -> 0.022538 ms（2.884x）。
• GitHub pre-run-check 成功。全仓 pre-commit 失败是主线格式基线问题：CI 重写 227 个文件，与本 PR 的 5 个文件交集为 0。

因此按源码、静态门禁、SM70 编译、V100 正确性与微基准证据批准合并。