[FSDP] Clean missing and unexpected keys (#120600)
Currently, when loading w/strict=False or w/strict=True and looking at
error message, FQNs are garbled w/FSDP details such as "_fsdp_wrapped_module".
This makes it tricky for upstream applications to validate the expected set of
keys are missing / unexpected (for example with PEFT where state_dict is loaded
non-strict), and makes error message more complicated w/FSDP details.
This PR cleans those prefixes by using `clean_tensor_name` in FSDP's existing
post load_state_dict hooks. Currently, only full_state_dict impl is tested, can test the rest of the impls as follow up work.
Differential Revision: [D54182472](https://our.internmc.facebook.com/intern/diff/D54182472/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120600
Approved by: https://github.com/XilunWu, https://github.com/fegin