Add system prompt to in-house judges

[martinscooper]

Changes:

• Criteria's prediction_field was renamed to to_evaluate_field .
• DirectInstance and PairwiseInstance models were removed and unified under the Instance model.
• Instance now just holds a fields attribute. context, response and responses fields were removed.
• The logic behind how the text to be evaluated and how the context is evaluated was re-designed. Now, the criteria defines the role of each of the instance fields.
• Lanchain usage was heavily reduced and replaced by custom logic and the dependency will be removed soon.
• EvalAssist in-house judges prompts were changes: the prompts are now using system prompts and the message format.
• Lanchain's output fixer were replaced with custom logic.
• Synthetic instance generation was improved.
• More tests were added.