🏗️🌊 Support resolving non-rank-based metrics

[mberr]

• convert classification metrics into classes (parallel to rank-based metrics)
• add side/prediction target support for classification metrics
• introduce MetricKey base class, with ClassificationMetricKey and RankBasedMetricKey; those implement the conversion of strings into fully qualified metric keys (i.e., inferring defaults, and resolving synonyms and spelling variants)
• annotate base Evaluator with generic metric key type - this allows static code analysis to check whether we use the correct key with the correct evaluator / evaluation results.

Fix #1235

[mberr]

It might make sense to move the metric resolution into the evaluator class. This avoids, e.g., that non-rank-based metrics resolve a rank-type.