Adding Active Label Cleaning code

[melanibe]

This PR adds InnerEye-DataQuality a stand-alone subfolder that contains associated to our pre-print "Active label cleaning: Improving dataset quality under resource constraints".
In particular this subfolder provides the tools to:

1. Train noise robust models (co-teaching, ELR, SSL pretraining and finetuning capabilities)
2. Run the label cleaning simulation benchmark proposed in the above mentioned manuscript.
3. Run the model selection benchmark.
4. All the code related to our benchmark datasets CIFAR10H and our proposed NoisyChestXray benchmark.