Optimizing Multilingual LLMs via Federated Learning: A Study of Client Language Composition