Epic: Ensemble completeness — ExtraTrees classifier, sample weights, forest feature importance

[Mec-iS]

Tracking issue for ensemble work that shares one design surface (base_forest_*, tree internals). Do not start code before an RFC is agreed here.

Member issues

• feature request: extra-trees classifier and regressor #183 — ExtraTreesClassifier (ExtraTreesRegressor already exists)
• Feature request: sample weights support in RandomForest (and other tree-based models) #356 — sample weights for RandomForest (and other tree ensembles)
• Variable Importance in Random Forest Analysis #124 — variable importance exposed at forest level

Why grouped

All three touch the shared forest base and tree split logic. One design pass avoids three colliding PRs:

• Sample weights must flow through bootstrap sampling and split-criterion weighting.
• Forest-level importances should reuse DecisionTreeClassifier::compute_feature_importances (src/tree/decision_tree_classifier.rs) and add the missing regressor-tree variant.
• Classifier parity work (feature request: extra-trees classifier and regressor #183) should land after weights/importances so the new class uses them from day one.

Constraints (repo conventions)

• Pure Rust, no macros; zero-copy data access via iterator(...) / view traits.
• Numeric logic on numbers/ traits only.
• Serde: parameter defaults must match Default::default() (see Serde JSON defaults don't match Default::default() #175 lessons).
• Doctests + mod tests {} per module.

RFC requirements

Post a short RFC comment before any PR: API sketch, trait bounds, serde story, no_std impact for v0.7.

Acceptance criteria

• RFC agreed by a maintainer
• Members implemented or explicitly descoped
• CHANGELOG updated