Thin N GEMM kernels

[Isalia20]

GEMMs with small count of N and large M are bound by reading A. On main, the kernel uses 64x128 NAX tiles which waste most of the columns. For the kernel added in this PR, each simdgroup computes a 32x32 block of output matrix.

A matrix tile stays in registers and B matrix is read from device memory with mpp. K is split across simdgroups per output block, if M is <2048 then K is split across 4 simdgroups in quarters of K. If M is >=2048 then K is split in 2 main parts with 4 simdgroups while tile size used to write to output being bigger, so 4 simdgroups cover 64 columns instead of 32 columns

bf16 perf on M5 Pro with macOS 27:

Model, layer	M x N x K	main	branch	speedup
Qwen3.8-27B in_proj_a/b	2048 x 48 x 5120	106.1	65.1	1.63x
GLM-5.3 indexer weights_proj	2048 x 32 x 6144	122.6	65.1	1.88x
gpt-oss-20b router, bias	2048 x 32 x 2880	48.0	33.2	1.45x
Wan2.2-T2V-A14B proj_out, bias	32760 x 64 x 5120	1715.7	1324.9	1.29x

I don't have base m5 so would be nice to benchmark for it as well (M5 base is also included in the dispatch for this kernel).

• AI usage disclosure:

AI was used to port kernels from my repo https://github.com/Isalia20/metalBLAS to here.

[zcbenz]

This is a really nice optimization, thanks!