Remove composable API's fully_shard from torchtnt example and test (#944)

Summary:
Pull Request resolved: #944

The fully_shard name is now used by FSDP2 (torch.distributed._composable.fsdp.fully_shard) and the Composable API's fully_shard (torch.distributed._composable.fully_shard) is being deprecated. Therefore, we want to remove torch.distributed._composable.fully_shard from torchtnt as well.
Deprecation message from PyTorch:
https://github.com/pytorch/pytorch/blob/main/torch/distributed/_composable/fully_shard.py#L41-L48

Reviewed By: fegin

Differential Revision: D65702749

fbshipit-source-id: a755fe4f0c7800184d62d958466445e313ceb796

Author: wz337

tests/utils/test_prepare_module_gpu.py

@@ -9,8 +9,6 @@
 import unittest
 
 import torch
-
-from torch.distributed._composable import fully_shard
 from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
 from torch.distributed.fsdp.fully_sharded_data_parallel import MixedPrecision
 from torch.nn.parallel import DistributedDataParallel as DDP
@@ -93,7 +91,7 @@ def _test_is_fsdp_module() -> None:
         model = FSDP(torch.nn.Linear(1, 1, device=device))
         assert _is_fsdp_module(model)
         model = torch.nn.Linear(1, 1, device=device)
-        fully_shard(model)
+        model = FSDP(model)
         assert _is_fsdp_module(model)
 
     @skip_if_not_distributed