Self-Supervised Multimodal Representation Learning for Neuroimaging