Tuning language models as training data generators for augmentation-enhanced few-shot learning