Representations as Language: An Information-Theoretic Framework for Interpretability