FedRW: Efficient Privacy-Preserving Data Reweighting for Enhancing Federated Learning of Language Models