fix(qwen36): restore MoE device-tilemap (MOE_GPU) correctness — masked down-dequant no-op

[James-CUDA]

Summary

Correctness fix (not a speed change). SPARKINFER_PREFILL_MOE_GPU=1 — the Qwen3.6 device-tilemap prefill path that skips the per-layer D2H counts sync — is silently broken on main: it produces garbage output.

Root cause: the MoE down projection's masked Q→int8 dequant runs with cols = moe_ffn = 512, which the vectorized fast-mask kernel declines (it only accepts cols that are a multiple of 1024). launch_gguf_dequant_rows_i8_mask ignored that false return, so the int8 down weights were never written and the down GEMM ran on stale/garbage data. This is the #586 corruption that forced the path default-off.

Fix: launch_gguf_dequant_rows_i8_mask now falls back to a correct masked slow kernel (single-pass, register-resident, identical dequant/amax/round to the host deq_rows_i8_kernel; dead experts exit via the counts mask) when the fast path declines. The default host-tilemap path is untouched, and SPARKINFER_PREFILL_MOE_GPU stays default OFF (the device path is not currently a prefill speedup — it must dequant all experts per group rather than the host path's live-only coalesced runs — so this fixes the opt-in flag rather than changing perf).

Proof (correctness, not speed)

Not a speedup PR — no bench.sh before/after and the RTX-5090 speedup box is intentionally left unticked. Evidence is the prefill fidelity gate, on RTX 5090 / Qwen3.6-35B-A3B-UD-Q4_K_M, qwen3_gguf_prefill_check <model> 512 16 with SPARKINFER_PREFILL_MOE_GPU=1:

	TOP1	KL
before (main)	0.375 (6/16)	0.349
after (this PR)	0.9375 (15/16)	0.010

The host-tilemap default path (MOE_GPU=0) is unchanged (TOP1 1.0 @512, byte-identical to before).

• Tested on RTX 5090 (sm_120) — correctness only; no speed claim

# main (MOE_GPU=1):   TOP1 6/16 0.3750   KL 0.34935
# this PR (MOE_GPU=1): TOP1 15/16 0.9375  KL 0.00974

[github-actions]

Closed — RTX 5090 checkbox not ticked

This PR was auto-closed because the template includes Tested on RTX 5090 as - [ ] (unchecked). Evaluation is opt-in — tick the box only after a real 5090 run.

To submit for review:

1. Edit this PR description (you can edit while closed): change to - [x] Tested on RTX 5090
2. Fill the decode and/or prefill before → after tables with real bench/scripts/bench.sh numbers showing improvement
3. Reopen this PR

If this PR does not need GPU eval (e.g. docs-only), remove the proof-of-speedup section from the description instead of leaving an unchecked box.

CONTRIBUTING.md

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