1+1>2: Can Large Language Models Serve as Cross-Lingual Knowledge Aggregators?