Interpretable and Fair Mechanisms for Abstaining Classifiers