Interpretability Illusions in the Generalization of Simplified Models