Splitting the mixers' projections with the Neural Engine #690

On an M4 Pro, the FFN split from #260 leaves the ANE waiting while the GPU runs the mixers' input and output projections, which are most of the rest of each layer. I extended the split to those projections, calibrated per Mac the same way as the FFN. At 14K and 32K tokens it is 1.62× and 1.53× faster than main on Qwen3.8-27B (MLX 4-bit). Decode is unchanged, and upstream's regression check passes. I'm opening a PR with the details and would appreciate any feedback on the approach.