[ML] Implement AucRoc metric for classification

[przemekwitek]

This PR implements AucRoc metric for evaluating classification.
It achieves it the following way:

• adds new field (top_classes_field) to the Classification evaluation object
• adds new metric (AucRoc) under classification package
• adds new method (EvaluationMetric::getRequiredFields) by which each metric may request fields it needs for its calculation

It also introduces a few refactorings:

• field names interesting to the evaluation flow are now stored in the EvaluationFields object (reused by all the evaluation types)
• EvaluationMetric::aggs now accepts the new EvaluationFields object instead of the individual field names
• calculation code is shared between classification.AucRoc and outlierdetection.AucRoc classes using a base class (AbstractAucRoc)

The newly added fields are never indexed (persisted) so they don't require mappings.

Relates #62160

[przemekwitek]

run elasticsearch-ci/packaging-sample-windows

[przemekwitek]

LGTM The only open question left is how we deal with docs whose actual field is the class in question, the point that Tom raised.

I'll discuss the necessary changes and push them in a separate PR as this one has grown big already.