Augment with Care: Contrastive Learning for the Boolean Satisfiability Problem