The more aggressive matmul kernel lost to the register budget

The WebGPU matmul sweep started with a naive kernel, then added 16 by 16 workgroup tiling and a 4 by...