Fix bugs of loading model checkpoints as HuggingFace LM

[qingyangliu0065]

Description

This PR fixed errors in loading HuggingFace model checkpoints as HuggingFaceLM objects. The init function of HuggingFaceLM now supports two ways of configuration:

• input a model name corresponding to a pre-trained model in HF model hub, and optionally the tokenizer name (if none, we will use the default one with the model name)
• input a path to the model checkpoint together with the tokenizer path (this could also be the pre-trained model name in most cases where tokenizers are not modified during training)

[neubig]

Looks good, thanks!