[ROCm] Add int4 support (pytorch#129710)

Add AMD support for int4 kernel using mfma_f32_16x16x16bf16 instruction.
Only supports CDNA2 and CDNA3 gpus for now.
Fixes pytorch#124699

Co-authored-by: Nikita Shulga <[email protected]>
Pull Request resolved: pytorch#129710
Approved by: https://github.com/malfet

Author: 2 people