Tuning Language Models as Training Data Generators for Augmentation-Enhanced Few-Shot Learning