[FSDP] Refactor optimizer in backward (#104813)
1) Use zero_grad(set_to_none=True) to set grad to None, 2) call
prepare_grad_for_optim() before call to .step, 3) use
_reset_flat_param_grad_info to set flat param gradient back to None. These
changes should just be refactors and equivalent to how gradient memory was
managed before.
Differential Revision: [D47310761](https://our.internmc.facebook.com/intern/diff/D47310761/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/104813
Approved by: https://github.com/awgu
