fix(gemv): instantiate mmvq_q4k_rows for K=6656 (Muse Glimmer)

[Tet-9]

What

The batched-row Q4_K launchers (launch_mmvq_q4k_rows / launch_mmvq_rows_f32 in
kernels/csrc/cuda/gemm/gemv.cu) already supported K=2048/4096/5120/6144 but not
K=6656 — Muse Glimmer's hidden size — even though the single-row launch_mmvq_q4k
already handles K=6656 via si_mmvq_q4k_kfixed_kernel<__nv_bfloat16, 26>.

Any batched verify-path call (e.g. speculative decode, M>1) against Muse Glimmer
hit the unsupported-K guard and either silently declined or fell back to N
serial single-row calls instead of the fused batched path.

Change

• Adds NSUPER=26 explicit instantiations of si_mmvq_q4k_rows_exact_kernel
  for {__nv_bfloat16, float} × {MMAX=6, MMAX=8}, matching the existing
  pattern for K=2048/4096.
• Extends the K guard and dispatch branch in launch_mmvq_q4k_rows to cover
  K=6656.
• launch_mmvq_rows_f32 already had a K=6656 guard entry but its fallback
  dispatch macro silently mapped any unlisted K (including 6656) to KB=24 —
  fixed to dispatch KB=26 explicitly for K=6656.
• Adds a standalone correctness+timing check,
  runtime/examples/mmvq_q4k_rows_6656_check.cpp, isolating this kernel at
  K=6656 from the model loader and KV cache (same pattern as
  nvfp4_gemm_check.cpp).

Verification

Tested on RTX 5090 (sm_120):

• Batched output is byte-identical to N independent serial launch_mmvq_q4k
  calls at K=6656.
• 1.79x faster than the serial fallback (0.0312ms batched vs 0.0558ms serial
  for N=4 rows, MN=8192).
  K=6656 rows: batch 0.0312 ms, 4 serial calls 0.0558 ms, 1.79x
  [PASS] K=6656 batched-row Q4_K matches serial reference

Also re-ran the existing dflash_gdn_checkpoint_gpu_test (K=2048/4096 paths)
to confirm no regression — still passes.

[github-actions]

Closed — RTX 5090 checkbox not ticked

This PR touches runtime/ and was auto-closed because Tested on RTX 5090 is not ticked (- [x]). Runtime changes must show a real 5090 before/after benchmark to enter the eval queue.

To submit for review:

1. Edit this PR description (you can edit while closed): add or restore the proof-of-speedup section and tick - [x] Tested on RTX 5090
2. Fill the decode and/or prefill before → after tables with real bench/scripts/bench.sh numbers
3. Reopen this PR

Non-speed runtime fixes that should not be evaluated need a maintainer hold label.

CONTRIBUTING.md

_{Automated by rtx5090-required CI. Maintainers and hold PRs are exempt.}