[FEA] Does it supports quantization-matrix-mul?

[bianxuxuxu]

The mixed-dtype-gemm example supports upcasting from a narrower (fewer bits) to a wider (more bits) type, but I need a quantization gemm which is from a wider to a narrow type.
For example, fp16xfp8 mm, we need do fp16 quantized to fp8 firstly(fp16/quant_scale, the quant_scale is provided), then do the fp8xfp8 gemm. How to do this?