Learning to Scaffold: Optimizing Model Explanations for Teaching