perf(moe): GeGLU fused decode-MoE; wire gemma4 (+13%) (#268 step 3b)

[inureyes]

Summary

Step 3b of #268: extend the two-kernel fused decode-MoE to GeGLU experts and route gemma4's MoE decode through it. gemma4's experts use gelu-tanh-approx(gate) * up, not SwiGLU.

What changed

• The gate/up kernel gains an act template arg: 0 = SwiGLU (silu), 1 = GeGLU (gelu tanh approx, matching compiled_geglu_approx_activation).
• The shared two-kernel body is factored into run_fused_moe_two_kernel, exposed as two FFIs: fused_moe_expert_kernel (silu — unchanged for qwen3_moe/dots/qwen3_next) and the new fused_moe_geglu_kernel.
• SwitchGeGLU gains forward_fused_kernel; Experts::forward dispatches to it on single-token decode behind MLXCEL_FUSED_MOE.

Result (gemma-4-26b-a4b-it, M1 Ultra)

Decode 73.8 → 83.2 tok/s (+13%) — the largest fused-MoE win so far. gemma4 previously used the compiled SwitchGeGLU path (3 gather_qmm + geglu); the fused kernel reads each weight once across all cores and folds the activation + score weighting into the GEMV epilogues. Its small experts (Dff=704) make the MoE a meaningful decode fraction.

Validation

• Greedy temp-0 output byte-identical to the compiled-switch path (with the chat template) on gemma-4-26b-a4b-it across multiple prompts. (Raw no-chat-template prompts make this instruct model degenerate, so parity must be checked with the template.)
• 4-bit affine, single-token decode, off by default (MLXCEL_FUSED_MOE).
• cargo fmt + cargo clippy clean.

Refs #268.