Self-Training with Purpose Preserving Augmentation Improves Few-shot Generative Dialogue State Tracking