Debugging and Explaining Generalization Errors of High-performance Metric Learning Approaches: An Influence Function Based Perspective