Generalize HF datasets to a collection of HF datasets via hf_datasets

[chimezie]

To support the following dataset configuration:

hf_datasets: 
  - hf_dataset:
        name: "Open-Orca/OpenOrca"
        train_split: "train[:90%]"
        valid_split: "train[-10%:]"
        prompt_feature: "question"
        completion_feature: "response"
  - hf_dataset:
        name: "trl-lib/ultrafeedback_binarized"
        train_split: "train[:90%]"
        valid_split: "train[-10%:]"
        chat_feature: "chosen"

adding support for custom chat HF datasets (addressing #1088 and relevant for #513) and fixing (#1087)

[chimezie]

Superseded by #1103