Empowering Large Language Models for Textual Data Augmentation