Fix saving of generation_config for Llama-3

[eldarkurtic]

The existing version of HuggingFaceCheckpointer does not respect model's generation_config.json and during saving it initializes it from the config.json. In the case of Llama-3 model, there is a discrepancy with eos_token_id which is set to 128001 in config.json but to [128001, 128009] in generation_config.json. This means that Llama-3 models saved with HuggingFaceCheckpointer have "eos_token_id": 128001 instead of "eos_token_id": [128001, 128009],.

This creates problems when we want to use Llama-3 models produced with llm-foundry as they will most likely always generate text until the max number of tokens is exhausted instead of stopping at 128009 token.

[dakinggg]

Thanks for the fix! Please run pre-commit and then we can merge