[Kernel] Expand SM70 75T prefill to Q8192 and concurrent requests

[yangzhuxinyzx]

Purpose

The default SM70 75-TFLOP/s long-prefill path only admitted Q8000 with a single-request paged-gather gate and an 8000-token KV step. That kept normal Q8192 chunks and concurrent scheduler batches off the optimized path.

This change adds a native Q8192 FP16 Tensor Core / FP32 MMA-accumulation specialization, dispatches Q8001-Q8191 by causality-preserving leading padding, and removes the single-request admission gate. Each eligible request in a scheduler batch now gathers paged KV into the reusable dense workspace and enters the same architecture path. The admitted KV range is Q through 262144 with 32-token alignment; boundary testing showed that smaller prefix PV remainders remain finite but are numerically invalid, so those shapes keep the general fallback.

The two-request end-to-end run also exposed an existing E4M3 batched-XQA planner mismatch: the Python planner could choose partition 1024 although the native B2-B16 kernel accepts 64/128/256. Automatic batched E4M3 planning is now capped at 256, while incompatible explicit overrides fail clearly.

Test Plan

• Build the SM70 FA2 extension from source and verify the Q8000 and Q8192 operators are registered.
• Benchmark Q8192 at aligned 128K and 256K-class KV lengths with 30 warmups and 100 CUDA-event samples.
• Compare sampled rows against an FP32 attention reference and run overflow/stability inputs.
• Run one 256K cold TP4 request and two concurrent 128K cold TP4 requests with prefix caching disabled and E4M3 KV.
• Run the SM70 route-policy, decode-planner, and CUDA stability suites.
• Run all pre-commit hooks on the changed files.

Test Result

• Final FA2 SHA256: 9f55da1ae54d87b008cb3452e23e1e4b272dd66115d37620d63674f3841c0055.
• Q8192 operator throughput:
  • KV128000: 76.7010 TFLOP/s, relative L2 0.2387%.
  • KV128032: 75.6171 TFLOP/s, relative L2 0.2350%.
  • KV256000: 75.8294 TFLOP/s, relative L2 0.2607%.
  • KV262144: 75.5520 TFLOP/s, relative L2 0.2290%.
• All sampled outputs were finite. KV=Q+8/+16/+24 boundary probes exposed incorrect partial PV tiles and are now rejected; KV=Q and Q+32 passed.
• One cold 256000-token TP4 request: 96.9743 s TTFT, 2639.88 prompt tok/s, zero cached tokens, and the exact same 16 output token IDs as the Q8000/control run.
• Two concurrent cold 128000-token TP4 requests: 74.2707 s batch TTFT and 3446.85 aggregate prompt tok/s over 256000 input tokens. Both requests had zero cached tokens and returned the same complete answer. Every TP rank recorded 448 native Q8192 calls and 480 E4M3 bridge calls.
• pytest: 172 passed.
• Pre-commit: all hooks passed, including Ruff, clang-format, markdownlint, mypy, SPDX, and DCO sign-off.