Add precision, recall and AUROC scorers; fix nondeterministic weight fitting

[gAldeia]

New classification scorers

• precision, recall and roc_auc for binary and multiclass problems. For multiclass, precision and recall are macro averages, and AUROC is a macro one-vs-rest average.
• average_precision_score now also works for multiclass (macro one-vs-rest).
• The new scorers only drive selection, survival and the archive. Weights are still fit with log loss, and splits with gini impurity.
• Class weights are applied as sample weights, matching sklearn. Values agree with sklearn to within 1e-4.
• The scorer can be changed between partial_fit calls. partial_fit now also updates parameters_, so the new scorer is used by final model selection.

Fix: Floor/Ceil broke weight optimization

These nodes were flagged as weighted but left out of get_weights(). During fitting, evaluation still took a weight for them. That shifted every later weight to the wrong node and read past the end of the parameter array. As a result, fits depended on uninitialized memory, and runs with the same seed could differ. The fix is that only weighable node types take a weight during evaluation. test_fixed_seed_produces_identical_brush_runs now passes consistently.

Docs and tests

• C++ tests for the new metrics, plus a regression test for the weight bug.
• Python tests comparing brush's metrics with sklearn's, and covering scorer changes in partial_fit.

[gAldeia]

Looks good. Thanks