[ML] Implement AucRoc metric for classification evaluation.

[przemekwitek]

Currently, the ML Evaluate API exposes AucRoc metric for evaluating the results of the outlier detection analysis.
Such a metric would also be useful for evaluating (multiclass) classification analysis.

[przemekwitek]

I consider this done.
There is still one outstanding issue with providing user-friendly error message when the old destination index is encountered in Evaluate request. But this is just a small enhancement that should not block this issue from being closed.