Improving Fairness of Large Language Models in Multi-document Summarization