Contextual Vision Transformers for Robust Representation Learning