Add preprocessing module for evaluation datasets

For `boolq`, we'd need to modify each example to be in the format `f"Passage: {passage} Question: {question} Answer:` — and then appending "True" or "False" and evaluating likelihoods. Possibly via adding an f-string argument to the zero-shot evaluator.

Outstanding:
* If we do this, how should we track the formatting for reproducibility of results? Sometimes removing trailing spaces or changing capitalization makes a big difference in accuracy, but you'd need to know the exact prompting string to reproduce results