Pushing the limits of self-supervised ResNets: Can we outperform supervised learning without labels on ImageNet?