Add metrics for robustness issues

[eu9ene]

We already calculate them in the training pipeline based on augmented datasets. For examples on all upper case dataset. It would be great to add them to the final report too and compares against the cloud providers. This requires integrating the same methods we use in the training pipeline, so likely reusing the code from there. We can wait until we have W&B dashboards. Maybe they will be sufficient.