Diffusion probabilistic models enhance variational autoencoder for crystal structure generative modeling