perf(moe): two-kernel decode-MoE that beats gather_qmm (#268 step 2b) - #276

Summary

Step 2b of the fused decode-MoE kernel effort (#268). The 2a kernel (#275) was correctness-validated but ran one threadgroup per token (one GPU core), so it was ~88x slower than gather_qmm. This PR parallelizes it into a kernel that beats gather_qmm by ~3.5% on qwen3-30b-a3b, greedy output byte-identical.

What changed

Two steps, each benched on qwen3-30b-a3b (M1 Ultra), greedy byte-identical throughout:

1. SIMD-cooperative GEMV (one simdgroup per output row, 32 lanes stride the contraction dim + simd_sum): 0.54 → 21.6 tok/s (40x over 2a), but still 0.46x of gather_qmm. An occupancy sweep (redundant tiling to force K·R threadgroups) confirmed the kernel is occupancy-bound, not tuning-bound — and that a fully-fused single launch can't escape it: the gate/up → down dependency runs through the per-expert activation, pinning each expert to one threadgroup.

   threadgroups	8 (K)	16	32	64
   decode tok/s	21.6	24.9	26.4	27.4

2. Two-kernel non-redundant split: stage the swiglu activation in global memory across two dispatches (A: gate/up + swiglu → act_g[K, Dff]; B: down · score → partial[K, Din], summed over K). Every GEMV output row is now an independent simdgroup across all cores, each weight read exactly once: 49.0 vs 47.3 tok/s, +3.5%.

Validation

• Greedy temp-0 output byte-identical to the gather_qmm path on qwen3-30b-a3b (3 prompts) and qwen1.5-moe-a2.7b.
• qwen1.5-moe-a2.7b (~140 tok/s) is at parity (no regression); the small/fast MoE has less to gain.
• cargo fmt --check clean, cargo clippy --release --features metal,accelerate clean.

Scope / flags

Single-token decode, 4/8-bit affine only, off by default (MLXCEL_FUSED_MOE). MLXCEL_FUSED_MOE_SGY tunes simdgroups per threadgroup (default 8). The single-launch kernel and its diagnostic SG/R knobs are removed in favor of the two-kernel implementation.

Next (not in this PR)

• Step 2c: 6-bit / mixed-bit unpack so dots.llm1 (6-bit down_proj, the 0.51x target) and nemotron-h MoE benefit instead of falling back.
• Broader-fleet validation and a decision on flipping MLXCEL_FUSED_MOE on by default.

Refs #268.