Boosting Generalization Performance in Model-Heterogeneous Federated Learning Using Variational Transposed Convolution