Have Your Cake and Eat It Too: Toward Efficient and Accurate Split Federated Learning