Text this: Safe Semi-Supervised Contrastive Learning Using In-Distribution Data as Positive Examples