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