Unity by Diversity: Improved Representation Learning in Multimodal VAEs