Customizing Language Model Responses with Contrastive In-Context Learning