Towards Energy-Aware Federated Learning via MARL: A Dual-Selection Approach for Model and Client