Few-Shot Adaptation for Parsing Contextual Utterances with LLMs