A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics