DemystifyingContrastiveSelf-SupervisedLearning: Invariances,AugmentationsandDatasetBiases