Improving Multi-Label Contrastive Learning by Leveraging Label Distribution