MAINT: Crescendo and Score Refactor

[rlundeen2]

This was the result of walking through Crescendo and stepping through code. I believe it greatly simplifies the logic, and also greatly increases efficacy of Crescendo. Additionally, I refactored several scorers to simplify/reuse code.

Description

• Refactored Crescendo and Crescendo tests to be simpler
• The Crescendo failure to respond message contained the refusal at the top by mistake. This resulted in less accurate attacks
• Conversation history in Crescendo incorrectly rolled back on JSON errors (those conversations were never added to memory) resulting in corrupted conversations
• eval_judge_true_false_scorer was updated to be a scale scorer with a task, increasing eval_score accuracy
• Fixed refusal scorer so it takes in a task, increasing refusal scorer accuracy
• Rewrote scale scorer to be a bit more flexible with examples.
• Refactored scorers to abstract self-ask paradigms and increase code reuse
• Removed the requirement that PromptTemplates use every parameter

Tests and Documentation

Tests written. Crescendo Jupytext run; Running tree of attacks