Metal backend: port quantized GEMM from MLX (#17362)

Optimized 4-bit linear in the Metal backend for the Matrix-Matrix case
(M > 1), by porting the quantized GEMM shaders (QMM) from MLX

Author: manuelcandales