[shard] add ShardedTensor.cpu() (#74941)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/74941
Add ShardedTensor.cpu() API, we choose to add this first instead of `ShardedTensor.to()` bc the latter one is ambiguous and tricker (i.e. we need to figure out what's the meaning of ShardedTensor.to("cuda")).
Note that when moving ShardedTensor to CPU from GPU, user need to explicitly pass in a new process_group that is compatible with the device, otherwise we will error out.
ghstack-source-id: 153455401
Test Plan: test_sharded_tensor_cpu
Reviewed By: pritamdamania87
Differential Revision: D35238133
fbshipit-source-id: 71921942f9295384f4e0eec2bfcdcff63a3d047c
(cherry picked from commit a7d97f70f81648f0f44cd77b4b064b4967aaa2d5)
