Answer When Needed, Forget When Not: Language Models Pretend to Forget via In-Context Knowledge Unlearning