Adversarial Auto-Augment with Label Preservation: A Representation Learning Principle Guided Approach