feat(evaluation): add ROC AUC metric

[tachyonicClock]

Summary

• Bump the bundled MOA jar (invoke.yml) to the version that adds a prequential ROC AUC to BasicClassificationPerformanceEvaluator (F1 fix + log loss + evaluate prequential delayed regression Waikato/moa#332).
• Enable the new rocAUC option by default in ClassificationEvaluator, and expose it as roc_auc() on both ClassificationEvaluator and ClassificationWindowedEvaluator, so it's accessible from PrequentialResults the same way as accuracy, kappa, f1_score, etc.
• Add the roc_auc -> "ROC AUC (cumulative)" name mapping used to translate between CapyMOA and MOA metric names.

Note: MOA computes ROC AUC prequentially for binary classification only, reporting NaN for multi-class problems (see adaptive-machine-learning/backlog#23, which had been blocked on an alternative approach — we ended up simply adding ROC AUC to the evaluator instead).

Test plan

• pytest tests/test_evaluation.py passes (47 passed), including test_evaluation_api which generically checks every metric returned by metrics_header() is accessible through PrequentialResults.
• Manually verified results.cumulative.roc_auc(), results.roc_auc() (delegated), and results.windowed.roc_auc() all return sensible values on ElectricityTiny with HoeffdingTree.

Assisted-by: claude-code:claude-sonnet-5