[Kernel][SM70] Avoid NVFP4 MoE prefill input expansion

[yangzhuxinyzx]

Purpose

Stacked on #387. Remove the Qwen3.8 TP4 NVFP4 prefill input-expansion bottleneck without changing routing order, W2, activation, unpermute, decode, or output arithmetic.

The exact E512/K10 W13 route keeps the existing stable expert sort, derives original token-row indices in a lightweight metadata kernel, and uses TurboMind SM70 MatrixLayout.idxs instead of materializing [tokens * 10, 2560] FP16 rows.

Safety

• Restricted to exact TP4 Qwen3.8 W13: E512/K10, K2560/N320, group size 16, at least 128 tokens.
• VLLM_SM70_NVFP4_QWEN38_MOE_INDEXED_PREFILL=0 restores the old route.
• Implicit default falls back with an older extension; explicit =1 fails closed when either operator is missing.
• Draft remains stacked on [Kernel][SM70] Group QSA Page4 prefill #387 until grouped QSA lands.

Test Result

• SM70 _C and _moe_C incremental builds and schema imports pass.
• Focused suite: 15 passed, 5 skipped.
• Ruff, clang-format, typos, and git diff --check pass.
• Locked V100 8192-token operator chain: 6.026752 -> 4.235264 ms (1.423x, 1.791488 ms/layer/rank saved). All routing metadata and W13 output are bitwise equal.
• One-start TP4 V2/no-MTP endpoint versus [Kernel][SM70] Group QSA Page4 prefill #387:
  • 32K: 5998.65 -> 6507.10 tok/s (+8.47%)
  • 64K: 5777.43 -> 6241.48 tok/s (+8.03%)
  • 131K: 5450.92 -> 5871.47 tok/s (+7.71%)
  • 8K: 6394.74 -> 6970.80 tok/s (+9.01%)
• Arithmetic, Chinese, 32K, 64K, 131K, 8K, and profile token hashes match [Kernel][SM70] Group QSA Page4 prefill #387 exactly.

Full implementation and benchmark evidence are recorded in docs/design/sm70_v100_migration_control.md.