GRU: Mitigating the Trade-off between Unlearning and Retention for Large Language Models