Why do We Need Large Batchsizes in Contrastive Learning? A Gradient-Bias Perspective