[sharded_tensor] easier initialization for Shard (#66351)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/66351
This add the ability for use to just provide shard_offsets and optionally rank, to construct a local shard, instead of knowing there's a ShardedMetadata. Under the hood, we will construct the ShardedMetadata by inferring shard_lengths and device from the local tensor.
ghstack-source-id: 141742410
Test Plan: test_local_shards
Reviewed By: pritamdamania87
Differential Revision: D31519919
fbshipit-source-id: 8f3b4682ffc74b79b41076f3f4b832f4cacda49d
