perf(moe): 6-bit/mixed-bit fused decode-MoE; wire dots.llm1 (#268 step 2c)

[inureyes]

Summary

Step 2c of #268: extend the two-kernel fused decode-MoE (#276) to 6-bit and mixed bit widths, and route dots.llm1's MoE decode through it (it fell back before, being 4/6-bit mixed).

What changed

The two-kernel split makes mixed widths natural — kernel A (gate/up + swiglu) and kernel B (down) each get their own bits template arg, so dots.llm1 (gate/up 4-bit, down 6-bit) needs only a 6-bit path in kernel B.

6-bit is not a power of 2: MLX packs 4 weights into 3 bytes (pack_factor 4, bytes_per_pack 3). The lane reads the down row as bytes and reconstructs the four 6-bit values with the bit layout taken from MLX quantized.h qdot:

v0 = b0 & 0x3f
v1 = (b0 >> 6) | ((b1 & 0x0f) << 2)
v2 = (b1 >> 4) | ((b2 & 0x03) << 4)
v3 = b2 >> 2

• FFI fused_moe_expert_kernel now takes gu_bits (gate/up, 4/8) + d_bits (down, 4/6/8) instead of one bits; group_size stays shared.
• Rust guards relax: gate/up must match each other and be 4/8-bit; down may be 4/6/8-bit; 6-bit down requires Dff % 16 == 0 (whole uint32 columns).
• dots1.rs MoE forward dispatches to the fused kernel on single-token decode behind MLXCEL_FUSED_MOE, mirroring qwen3_moe.

Validation (M1 Ultra, greedy temp-0)

• dots.llm1 (the 6-bit path): output byte-identical to the SwitchGLU fallback. Decode 13.1 → 13.7 tok/s (+4.7%).
• qwen3-30b-a3b: byte-identical (no regression from the signature change).
• cargo fmt --check + cargo clippy clean.

Still off by default (MLXCEL_FUSED_MOE), single-token decode only.

Scope

nemotron-h's MoE runs through the separate C++ full-forward path (fused_moe_forward), not the Rust SwitchGLU, so it is not wired here — a possible follow-up.

Refs #268.