Privacy-Preserving In-Context Learning for Large Language Models