[matmul] Add naive mmt benchmark

[kuhar]

The performance on Pixel 6 is roughly the same as the i8 matmul w/ innerproduct.

I also tried the following optimizations but they did not help:

• Vectorizing accesses to the output matrix (along the N dimention)
• Using subgroup operations to reuse 'lhs' across threads
• Creating a local matrix for inputA and/or inputB