[FSDP()][13/N] Refactor unshard/reshard/grads (#87926)
This PR is not too complicated. We just move unshard/reshard/grads out to `_runtime_utils.py` and make them take `state: _State` instead of `self`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87926
Approved by: https://github.com/mrshenli
