fsdp refactoring (#2177)
* remove the redundant code post the torch 2.1 release
* make `use_orig_params=True` by default.
* fix `save_state` optimizer saving for fsdp and update the fsdp example
* quality
* fixing the utils and tests. Updating the docs
* bump up the minimum version for FSDP support.
* address comment
* rename fsdp model checkpointing variables
