Optimize Gemma 4 26B-A4B inference on Apple Silicon. Speed up the MLX runner, transforms, Metal kernels, or speculative decoder while preserving output fidelity.