Stabilized In-Context Learning with Pre-trained Language Models for Few Shot Dialogue State Tracking