In-Context Unlearning: Language Models as Few Shot Unlearners