Towards Efficient and Effective Unlearning of Large Language Models for Recommendation