Fast Metal kernels in llama.cpp now cover all 26 Mac GPU weight formats

The contributor says a newly merged llama.cpp PR extends fast Metal kernels from 10 Mac GPU weight formats to all 26, including BF16, MXFP4, Q2_K, Q3_K and IQ quants. They say the earlier kernels sped up speculative decoding for 10 formats, while matrix multiplies on an M3 Ultra run up to 4.4 times faster when a model checks several tokens at once.