add mul_mat_id, and compile upstream's kernels as they ship

vendor only the files the build reaches

Add Metal backend; vendor whole llama.cpp trees

GGUF kernels: dequantize + fused gemv over packed blocks, 9 CUDA variants