cuda: NVFP4 experts' prompt rows take the warp's row tile count as a constant

[Bizuayeu]

What this changes

The staged NVFP4 expert prompt kernel reads its row-tile count at run time inside the mma chain; this makes it a
template constant with one instantiation per count, so the branch leaves the chain. The sums keep their order and
every output row keeps its bytes. #764 (skip on the same prompt entries) touches the same three call sites; whichever
lands second takes the other's lines with a small rebase.

Receipt

• Environment: 1.0.9 (c36b536d) with this commit for the tests, on one NVIDIA GB10 (sm_121, CUDA 13.3). The
  exactness and timings are the same commit on 1.0.8 (0efbd447): the kernels on that GB10, the prefill through
  the GLM-5.3-Flash CUDA family (glm: GLM-5.3-Flash (NVIDIA's NVFP4 checkpoint) as a CUDA family at TP over hosts #804, nvidia/GLM-5.3-Flash-NVFP4 at 423acf37) at TP=2 on two (driver
  580.173.02, CUDA 13.0, Linux 6.17.0-1032-nvidia).
• Exactness: the old and new kernels' outputs (swiglu, down, relu²) hash equal at 1 to 2,048 rows in 11 row counts on
  random inputs at GLM-5.3-Flash's per-rank widths (1,024 x 4,096, 17 experts, 9 slots); tf-cuda-test quant-ref
  4 passed. Through the family at TP=2 the 8k and 32k prompt checks gave the same ids with and without it. On
  1.0.9, tf-cuda-test nvfp4-experts against the Python engine's recorded bytes (packing, plan, up and down, at its
  default shape and a GLM-like one) passes with and without it.
• Prompt speed: not from tools/prefill_cold.py (1.0.0 removed it; the family is not upstream yet). Kernels at
  2,048 rows (ncu, median of 10): swiglu 5,565 → 3,912 us (29.7% shorter), down 2,500 → 2,116 (15.4%), relu²
  4,673 → 3,810 (18.5%). The family's prefill at TP=2 (median of 2): about 2k tokens 1.540 → 1.480 s, 8,192
  tokens 5.610 → 5.290 s (5.7% shorter), 32,768 tokens 21.855 → 20.705 s (5.3% shorter).
• Decode speed: unchanged by construction (the decode kernels are not touched).
• Tests run (empty cache, this commit alone on 1.0.9): zig build test-cuda-host 429/429, zig build test 644/649
  (5 skipped), tf-cuda-test quant-ref 4 passed, zig fmt.
• Not run, and why: Nemotron's NVFP4 prompt path (no checkpoint here); Metal (not touched; no Mac here).

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off.
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo.
• No precision traded for speed. If the bits change, the description says which sums change and why.
• Prompt processing is no slower than the last release.
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (No new test: every
  output keeps its bytes, which quant-ref, nvfp4-experts and the comparison above check.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.

[Bizuayeu]

Updated the description: on 1.0.9, tf-cuda-test nvfp4-experts against the Python engine's recorded bytes passes with
and without this commit, and the checklist's test item is ticked with its reason (no new test; the outputs keep their
bytes).

[ashhart]

Thank you, @Bizuayeu. Nemotron's NVFP4 expert prompt rows take the warp's row tile count as a compile-time constant, so one mma body runs each count. The store mask is still m0 < cnt.

On a GB10, teacher-forced rows are bit-equal to the base, 16,109 of 16,109. Served replies are equal, 17 of 17. Drafted replies equal "draft": false, 17 of 17. Cold prefill is 4.1% faster at 2k tokens, 7.5% at 8k, 6.9% at 16k, 6.6% at 32k, and 5.3% at 64k.

Your commit is in 1.0.10 under your name (4e63c6528). The patch matches your head (4ed879dc3). The CHANGELOG line is 343e64398. Your head 4ed879dc3 is merged at f2433079b with no change to the tree. We wrapped prompt_body's parameter lines (46ebd305a).