Personalized Federated Learning with Theoretical Guarantees: A Model-Agnostic Meta-Learning Approach