Enhancing the Generalization of Personalized Federated Learning with Multi-head Model and Ensemble Voting