Propagate name_or_path through HF Checkpointer

[snarayan21]

Makes sure to propagate model name_or_path through HF checkpointer (similar to how we do for generation config).

Note that since PretrainedConfig always has the attr name_or_path, we don't need to check hasattr. All instances of PreTrainedModel will have name_or_path (which defaults to "").

Note that if the pretrained_model_name_or_path is a local path then name_or_path will also be a local path. And so when saving out LoRA adapters, for example, the adapter_config.json will have base_model_name_or_path as the local path. This PR does not solve this problem. IMO we should add a new optional arg to ComposerHFCausalLM that's the true model name and propagate that through model transforms & HF checkpointer instead of name_or_path.

[snarayan21]

not needed rn