…t don't need a model (pytorch#1349)

These changes add a little complexity with the lazy and local imports, but
they also greatly improve the CLI's response for --help, list, and where.

Changes:

* Move `import torch` into function bodies that need them
* Use `importlib.metadata.version` to check the torch version rather than
  torch.__version__
* Switch from using torch.inference_mode as a decorator to using it as a
  context manager.
  * I also removed it from convert_hf_checkpoint_to_tune since that does
    not use torch at all
* In build_utils, wrap the dtype values in lambdas so they're lazily
  fetched.

pytorch/torchchat#1347
Branch: FasterCli-1347

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>