Closing the Generalization Gap in Parameter-efficient Federated Edge Learning