cuda: fix an experts-rows test race and measure the ModelOpt checkpoint on one GB10

[balancedscorpion]

For ashhart#952 (ashhart#952). Two commits on top of beec1aad6, independent of each other, so either can be
taken alone.

What this changes

• test(cuda): clear experts-rows buffers on the kernels' stream. Rig.zeros in
  zig/tests/cuda/experts_rows_tests.zig cleared each buffer with fill8(0, null), which is cuMemsetD8 on the
  default stream. The kernels run on a non-blocking stream that does not wait for it, so on sm_121 the clear of the
  one-row reference's output sometimes landed after the decode kernel wrote it, and tf-cuda-test experts-rows
  failed: FAIL experts_rows SwiGLU rows: first difference at byte 0, 76800 of 76800 words differ. The change clears
  on the rows' stream. One line, test only.
• docs: measure the ModelOpt FP4-Local checkpoint on one GB10. A subsection, "The ModelOpt checkpoint on one
  GB10", in docs/recipes/flash-next-cuda.md: decode, concurrency, cold first tokens to 131,072 tokens, the
  exactness and seam checks, and what a 128 GB machine admits. All measured with beec1aad6 built unpatched. Docs
  only.

Receipt

• Environment:
  • TensorFold beec1aad6c967f6e32aa14a3fd21ca8358710279, built with -Dnvcc -Dsm=121 -Doptimize=fast, nvcc 13.3,
    zig 0.17.0, served with TENSORFOLD_CUDA_KERNELS=native.
  • One NVIDIA GB10 (sm_121, 128 GB unified), driver 580.178.04, Ubuntu 24.04.5 aarch64.
  • nvidia/Qwen3.8-Flash-Next-FP4-Local revision b9ce18a6887e18c089fdea8509a33e63429a3300. No MLX or PyTorch.
• Exactness, served code unchanged by this PR (cuda: add experimental native Flash Next family and local tensor parallelism ashhart/TensorFold#952 as shipped, greedy, --prompt-cache-gib 0):
  • bench_concurrent.py --alone --serial --levels 1,4,8 --tokens 256 --reps 3, three runs: concurrent equals solo
    78/78 and drafted equals "draft": false 6/6 in each.
  • A cancelled reply, and a reply after a restart, equal the fresh reply's token_sha256. A second conversation
    turn resumed from the prompt cache (--no-drafts) equals the same turn computed without it.
  • flashnext-fidelity-single against -serial at 7 fixed histories of 20 to 6,496 tokens: 7/7 byte-equal.
• Decode speed before and after: n/a, no served code changes. The recipe's numbers are cuda: add experimental native Flash Next family and local tensor parallelism ashhart/TensorFold#952 as shipped: 64.1-64.6 /
  40.7-41.1 tok/s (fibonacci-raw / gpu-chat-no-think, one stream).
• Prompt speed before and after: n/a, no change reaches prompt processing. The recipe's cold first tokens at
  2k / 8k / 32k / 64k / 128k: 1.83 / 7.11 / 29.2 / 60.1 / 131.7 s (bench_cache.py --require-cold, exact arrays).
• Tests run:
  • tf-cuda-test experts-rows, thirty runs each, alternated, same build flags: before 8/30 failed, after 0/30.
  • tf-cuda-test kernel-contract on sm_121: SUMMARY 32 kernels: 32 pass gates 1-4, 0 fail, 0 known findings; 5 controls: 5 red as planted, 0 missed; 0 limit problems.
  • CI's No GPU job: zig build test 104/104 steps, 1,483/1,490 passed, 7 skipped (libjpeg-turbo absent);
    test-kernel-contract, test-golden and fuzz-cortheon pass.
  • test-cuda-host test-flashnext-contract-host for sm_121: 204/204.
  • lean_check.py: 0 new. zig fmt --check build.zig zig: clean.
• Not run, and why:
  • Metal, SM 80/86/89/120 and tensor parallelism: no hardware.
  • Six other test files pair fill8(0, null) with non-blocking streams (exlr8_tests, bench, oracle_tests,
    exlr8_dense_tests, libs_tests, runtime_tests). I have not seen them fail and did not change or check them.
  • A 262,144-token cold prompt: the window fitted a stream in one start of four here.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off. (No served
  code changes.)
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo.
• No precision traded for speed. No bits change.
• A new or changed GPU kernel is registered with the kernel contract. (No kernel changes; the sm_121 SUMMARY
  line is above.)
• Prompt processing is no slower than the last release. (Untouched.)
• 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. (An existing GPU
  test: 8/30 failed before, 0/30 after.)
• 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.