[FSDP][optim_state_dict] Returns the initial states of the empty parameters for KeyedOptimizer/NamedOptimizer (#94130)
KeyedOptimizer and NamedOptimizer expect the states exist in the state_dict when `load_state_dict` is called even if the corresponding parameters are empty (size == 0). This PR adds the support to make KeyedOptimizer work with `use_orig_params=True`.
Differential Revision: [D43019458](https://our.internmc.facebook.com/intern/diff/D43019458/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/94130
Approved by: https://github.com/rohan-varma
