qwen36 MoE prefill: fix MOE_GPU device-tilemap correctness (#586) + int8 projection default

[James-CUDA]

Summary

Two independent Qwen3.6 MoE-prefill changes:

1. fix(qwen36): restore MoE device-tilemap (MOE_GPU) correctness. SPARKINFER_PREFILL_MOE_GPU=1 (device tilemap, skips the per-layer D2H counts sync) is silently broken on main: the down projection's masked Q→int8 dequant runs with cols=moe_ffn=512, which the vectorized fast-mask kernel declines (cols not a multiple of 1024), and 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 data (the #586 corruption). Fix: fall back to a correct masked slow kernel (single-pass, register-resident, same dequant/amax/round as the host deq_rows_i8_kernel; dead experts exit via the counts mask). The default host-tilemap path is untouched and MOE_GPU stays default OFF (the device path isn't a prefill speedup — it dequants all experts per group vs the host path's live-only coalesced runs — so this fixes the opt-in flag, not perf).

prefill_check … 512 16, MOE_GPU=1	TOP1	KL
before (main)	0.375	0.349
after	0.9375	0.010

2. perf(qwen36): int8 tensor-core projections default-on for the MoE-hybrid prefill. Flips the use_i8 default for the Qwen3.6-256E MoE path (attn/GDN/router-feeding projections), scoped to Qwen3.6 by the existing n_experts==256 guard (Qwen3-30B is 128E, untouched). Scored accuracy is provably unchanged (qwen3_gguf_score teacher-forces via the bf16 decode forward_token path, which SPARKINFER_PREFILL_I8 never reaches — verified byte-identical PPL/ARGMATCH on/off); prefill fidelity holds (TOP1 0.875 @512 / 0.94 @4k, bar 0.80).

⚠️ Honesty note on speed: this was measured +10.6% pp @512 / +18% @4k vs the bf16 baseline, but main has since defaulted MoE projections to fp8 (#595), which captures the same win a different way. Against current main, the int8 flip is ~+0.6% @512 (fp8 and int8 are within noise). Kept in this PR per maintainer request; the standalone value here is the correctness fix above. SPARKINFER_PREFILL_I8=0 restores bf16.

Proof of speedup

• Tested on RTX 5090 (sm_120) — int8 flip is within noise vs the current fp8 default; no end-to-end speed claim against current main. The primary contribution is the correctness fix.

Prefill pp tok/s — Qwen3.6-35B-A3B-UD-Q4_K_M, RTX 5090 (int8 default vs SPARKINFER_PREFILL_I8=0 fp8, current main):

ctx	fp8 (current main)	int8 (this PR)
512	5726	5745
4096	16410	16403

# correctness (the real point of this PR), MOE_GPU=1, prefill_check 512 16:
# main:     TOP1 6/16 0.3750   KL 0.34935
# this PR:  TOP1 15/16 0.9375  KL 0.00974

[skyrocket2026]

⏳ Needs a benchmark to be evaluated

You ticked Tested on RTX 5090 but neither the decode nor the Qwen3.5 prefill before → after table shows a real claimed improvement. The on-device eval won't run until at least one does.

Decode (end-to-end, not an isolated-kernel microbench):

bench/scripts/bench.sh --download            # baseline (before)
bench/scripts/bench.sh --download            # your branch (after)

Prefill pp (Qwythos — report your best of 4k/32k/64k from the prefill pp line in bench output):

bench/scripts/bench.sh --download --ctx 4096   # try 4096 / 32768 / 65536 / 131072
bench/scripts/bench.sh --download --ctx 4096   # your branch

Then the bot greenlights it on the next poll and evaluates it on an RTX 5090.