Eurus prime data

[AlexPiche]

• Replace the test set of Eurus alexpiche/math_test_cleaned (which was built from eurus data) with HuggingFaceH4/MATH-500
• Get the prompt out of the Eurus dataset and into the config yaml file. \
• Add an argument to switch between sum and mean mean of the loss.
• Update docstring of extract_tape_training_samples (remove llm_call and strict)