Overfitting or perfect fitting? Risk bounds for classification and regression rules that interpolate