Federated Learning for Non-factorizable Models using Deep Generative Prior Approximations