Split up Simple Staging and Legacy Staging Logic (#1022)

Summary:
Pull Request resolved: #1022

Currently base_async_storage_writer is quite complicated and has a bunch of branching between simple_staging and legacy staging. We are looking to introduce yet another simple staging solution. So refactor as follows:

Last half, we introduced stager as an argument to DCP async_save.

1. Introduce LegacyStager and move legacy staging logic intop it
2. Introduce SimpleStager and move simple staging logic into it.
3. Based on JKs in checkpoint_dist_client initialize the right stager to use.
4. Remove Staging Logic From BaseAsyncStorageWriter

Differential Revision: D79434609

Author: Saiteja64

torchtnt/framework/callbacks/dcp_saver.py

@@ -26,6 +26,7 @@
     DefaultSavePlanner,
 )
 from torch.distributed.checkpoint.planner import LoadPlanner, SavePlanner
+from torch.distributed.checkpoint.staging import AsyncStager
 from torchtnt.framework.state import EntryPoint
 
 try:
@@ -161,6 +162,7 @@ def _checkpoint_impl(
         hook: str,
         planner: Optional[SavePlanner] = None,
         storage_writer: Optional[StorageWriter] = None,
+        stager: Optional[AsyncStager] = None,
     ) -> bool:
         if hook not in [
             "on_train_step_end",
@@ -199,6 +201,7 @@ def _checkpoint_impl(
                     process_group=self._process_group,
                     storage_writer=storage_writer,
                     planner=planner,
+                    async_stager=stager,
                 )
                 self._prev_snapshot = cast(Future, prev_snapshot)
                 if curr_snapshot_wait: