Connecting Interpretability and Robustness in Decision Trees through Separation