Generate and score responses (LLM-as-judge) with API-based models

[nouhadziri]

This PR does two things:

• API-mased model response generation: In addition to generating completions from HF and local models, the code now supports generating completions using API-based models (e.g., GPT-4). In a future PR I will address sampling responses from an ensemble of HF and API models.

• Response scoring with API models (aka LLM-as-judge): We now use API-based models to score responses based on the given prompt and scoring criteria. Note that this approach is different from prompting the model to rank multiple responses simultaneously. We need to address this in a separate PR and compare the results.

I have added a file prompt_templates.py which contains two types of prompts: i) generation and ii) judgment. This file also supports different skills, which will allow us to experiment with prompts for each skill added to Tulu.

[vwxyzjn]

LGTM. Thanks @nouhadziri!