Towards understanding neural collapse in supervised contrastive learning with the information bottleneck method