On Mutual Information in Contrastive Learning for Visual Representations