Learning from Training Dynamics: Identifying Mislabeled Data Beyond Manually Designed Features