[FSDP][optim_state_dict] Use all_gather to deal with uneven size tensors (#92991)
The current `_all_gather_optim_state` pads the uneven tensors which is not necessary as `all_gather` support the uneven tensors. This PR removes the padding logic.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/92991
Approved by: https://github.com/rohan-varma, https://github.com/awgu