y_true would be a label array (num_samples, ) or one hot matrix (num_samples, num_classes)
y_pred would be matrix of scores (num_samples, num_classes)

I think that we should be extending the scope to the ROC curve as well. An example where we do compute the different averages or handle multiclass is this example: https://scikit-learn.org/stable/auto_examples/model_selection/plot_roc.html#sphx-glr-auto-examples-model-selection-plot-roc-py

This would be good to have a simpler API to achieve the same results.

Without a scientific / statistical justification (e.g. literature), I'm hesitant to add such aggregations because, to me, it is not clear what such aggregations produce and how to interprete them. So my opposite proposal is to remove that part from the examples.

Hi! I have a PR for this feature now, all tests passing etc! Sounds good to include for ROC curve in the future, although I limited the scope of this particular PR to PR (!) only :)