Unlocking the Value of Decentralized Data: A Federated Dual Learning Approach for Model Aggregation