Reducing and Exploiting Data Augmentation Noise through Meta Reweighting Contrastive Learning for Text Classification