feat(pt): support CPU parallel training with PT (#4224)

Fix #4132.

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

- **New Features**
- Enhanced backend selection for distributed training, allowing for
flexible use of NCCL or Gloo based on availability.
  
- **Bug Fixes**
	- Corrected indentation for improved code clarity.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

deepmd/pt/entrypoints/main.py

@@ -105,8 +105,7 @@ def get_trainer(
     local_rank = os.environ.get("LOCAL_RANK")
     if local_rank is not None:
         local_rank = int(local_rank)
-        assert dist.is_nccl_available()
-        dist.init_process_group(backend="nccl")
+        dist.init_process_group(backend="cuda:nccl,cpu:gloo")
 
     def prepare_trainer_input_single(
         model_params_single, data_dict_single, rank=0, seed=None