Boost performance of MXFP4 quantization with inline PTX (#4694)

Summary:
Pull Request resolved: #4694

X-link: facebookresearch/FBGEMM#1720

- Add inline PTX to boost MXFP4 quantization kernel performance
- Fix MXFP4 scaling factor in grouped GEMM

Differential Revision: D80182398

Author: jiawenliu64