Forget to Know, Remember to Use: Context-Aware Unlearning for Large Language Models