[BE][FSDP] Enhance summon full params test (#79642)
Summary:
To resolve https://github.com/pytorch/pytorch/issues/74166, ensure
that when main FSDP API is CPU offloading, summon_full_params exit always
results in FSDP params residing on CPUs, not GPUs, as expected. Ensuring that
we call `_free_full_params_and_use_local_shard` in summon_full_params does
this.
Test Plan: CI
Differential Revision: D37184499
Pull Request resolved: https://github.com/pytorch/pytorch/pull/79642
Approved by: https://github.com/awgu
