Enabling stacked NVFP4 quantization kernels (#4458)

Summary:
Pull Request resolved: #4458

X-link: facebookresearch/FBGEMM#1518

Enable E2E under a modified setting where the input x is already concatenated and only m_sizes on GPU is available. The new design is compatible with cuda graph usage at the cost of a few extra kernels.

Reviewed By: q10, jiawenliu64

Differential Revision: D77420779

fbshipit-source-id: 891ab3f99356a279b608c494bb9eac7a8745f084

Author: Tianyu Liang