CSS: Combining Self-training and Self-supervised Learning for Few-shot Dialogue State Tracking