Add thin M GEMM dispatch

[Isalia20]

Followup to #4640. Adds thin M GEMM dispatche logic and the tests which I left out in the previous PR accidentally. Speed on bf16 M5 Pro:

Model, layer	M x N x K	main	branch	speedup
Qwen3.8-27B mlp.gate_proj	32 x 17408 x 5120	984.3	694.1	1.42x
Qwen3.8-27B mlp.down_proj	32 x 5120 x 17408	1039.4	735.4	1.41x
Llama-3.1-8B mlp.gate_proj	64 x 14336 x 4096	717.7	461.3	1.56x
Llama-3.1-8B mlp.down_proj	64 x 4096 x 14336	721.7	543.2	1.33x
Qwen3-8B mlp.gate_proj	16 x 12288 x 4096	622.4	395.0	1.58x
gemma-4-31B mlp.gate_proj	48 x 21504 x 5376	1314.4	917.6	1.43x
Qwen-Image text encoder mlp.gate_proj	40 x 18944 x 3584	792.0	541.2	1.46x
Qwen-Image text encoder mlp.down_proj	40 x 3584 x 18944	832.7	614.3	1.36x

• AI usage disclosure:
  N/A