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

vendor only the files the build reaches

drop the cuda kernel: metal is the only backend built for now

name the package after the repo it publishes to

point the hub upload at the renamed repo

trim the metal shader to the dispatched kernels

full tree: sources, vendored ggml, and all built variants

Add Metal backend; vendor whole llama.cpp trees

GGUF kernels: dequantize + fused gemv over packed blocks

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