[shard] fix named_params_with_sharded_tensor (#70228)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/70228
fix named_params_with_sharded_tensor impl, where `named_parameters` already loop the submodules recursively, so we shouldn't put it in the submodule loop.
ghstack-source-id: 146076471
Test Plan: Added more complicated test cases (that involves multiple submodules) to capture this issue.
Reviewed By: pritamdamania87
Differential Revision: D33251428
fbshipit-source-id: cf24ca7fbe4a5e485fedd2614d00cdea2898239e