The Power of Two: integrating deep diffusion models and variational autoencoders for single-cell transcriptomics analysis