[train v2] Fix trainer import deserialization when captured within a Ray tasks

[justinvyu]

Summary

Avoid imports from ray.train.torch in the TorchTrainer definition file since it causes circular import errors when captured in the scope of Ray tasks.

The following script errors at the moment with a circular import error when the TorchTrainer class gets deserialized on the task/actor that captured it. This is also an issue for the Tune + Train integration because the Tune function trainable gets run as a Ray actor task, which would capture the imported class.

# RAY_TRAIN_V2_ENABLED=1

from ray.train.torch import TorchTrainer

@ray.remote
def task():
    TorchTrainer

ray.get(task.remote())

Traceback (most recent call last):
  File "/home/ray/anaconda3/lib/python3.11/site-packages/ray/_private/function_manager.py", line 647, in _load_actor_class_from_gcs
    actor_class = pickle.loads(pickled_class)
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ray/anaconda3/lib/python3.11/site-packages/ray/train/v2/torch/torch_trainer.py", line 4, in <module>
    from ray.train.torch import TorchConfig
  File "/home/ray/anaconda3/lib/python3.11/site-packages/ray/train/torch/__init__.py", line 28, in <module>
    from ray.train.v2.torch.torch_trainer import TorchTrainer  # noqa: F811
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ImportError: cannot import name 'TorchTrainer' from partially initialized module 'ray.train.v2.torch.torch_trainer' (most likely due to a circular import) (/home/ray/anaconda3/lib/python3.11/site-packages/ray/train/v2/torch/torch_trainer.py)