`hf_access_token` only effective for loading gated datasets, not gated tokenizers

[James Chen]

Seems the hf_access_token is only effective when loading a gated dataset, not necessarily a gated tokenizer.

One can work around this via modifying preprocess_dataset() in _hf_data_processing.py to explicitly call huggingface_hub.login(token=config.hf_access_token if a token is present.