PERFEX: Classifier Performance Explanations for Trustworthy AI Systems