Statistics for Controllably Sparse Perturbations of Robust Classifiers for Explaining Predictions and Probing Learned Concepts