add DTensor to VLE and fix 2D sharding group in EBC (#2626)

Summary:

Adding DTensor state dict to VLE, this is identical to EBC DTensor path.

Future work of this diff is to consolidate the logic of EBC/PEA/VLE state dict into one parent class because of their significant similarities

A revert diff removed the 2D sharding logic in embedding bag collection mistakenly: D66800554, this diff adds it back in

Differential Revision: D65555595

Author: iamzainhuda

torchrec/distributed/embeddingbag.py

@@ -65,6 +65,7 @@
     QuantizedCommCodecs,
     ShardedTensor,
     ShardingEnv,
+    ShardingEnv2D,
     ShardingType,
     ShardMetadata,
 )
@@ -938,7 +939,11 @@ def _initialize_torch_state(self) -> None:  # noqa
                     ShardedTensor._init_from_local_shards(
                         local_shards,
                         self._name_to_table_size[table_name],
-                        process_group=self._env.process_group,
+                        process_group=(
+                            self._env.sharding_pg
+                            if isinstance(self._env, ShardingEnv2D)
+                            else self._env.process_group
+                        ),
                     )
                 )