Describir: Survey on the Role of Mechanistic Interpretability in Generative AI