Matmul() using PyTorch's MPs back end is faster than Apple's MLX