Open Problems in Mechanistic Interpretability