[Easy][FSDP] Clarify `_use_unsharded_grad_views` comment (#100359)
This is an easy follow-up to the previous PR to (1) clarify that `view` is the original parameter's gradient and (2) that after `reshard()` the gradient is on CPU only if offloading parameters.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100359
Approved by: https://github.com/rohan-varma
