Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability