Confidence-Aware Personalized Federated Learning via Variational Expectation Maximization