[FSDP] Update `ShardingStrategy` and `_free_full_params()` docs (#80894)
1. I messed up the comment for the post-backward `_free_full_params()` in https://github.com/pytorch/pytorch/pull/75901.
This removes the comment, which is not necessary, and instead adds an explanation in the `SHARD_GRAD_OP` comment itself.
2. This updates the overall `ShardingStrategy` documentation after the observation that `SHARD_GRAD_OP` did not specify that parameters are still sharded outside of computation.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/80894
Approved by: https://github.com/rohan-varma, https://github.com/zhaojuanmao