What Makes a " Good " Data Augmentation in Knowledge Distillation - A Statistical Perspective