Unlearning cleanup

[adamkarvonen]

Changes made:

• Move llm and sae to the llm_dtype, resulting in a ~5x speedup on Gemma-2-2B.
• Move hardcoded hyperparameters to the eval_config.
• Generate question_ids for the model if the question_ids don't exist
• Save a correctly formatted output filename for the graphing notebook
• Iterate over multiple retain_thresholds
• Cleanup unused functions
• Various minor fixes