Request for mixed precision (i8 and f16) gemm support

[sjw36]

From Harris:

There are LLMs in the zoo that have i8 weights but compute in f16. These can save memory and bandwidth by storing the parameters in lower precision.

For GEMM (and Conv2d) support rocMLIR should allow fusion of conversion functions (i.e. tosa.cast) on the inputs.

I will update with a link to these models.