Improving Multimodal Joint Variational Autoencoders through Normalizing Flows and Correlation Analysis