feat(ml): ensemble classifiers from cv

[junkha]

Summary

This PR introduces ensemble classifiers that combine fitted estimators produced by mother_cv for prediction and uncertainty quantification.

Changes

• Introduce CVEnsembleClassifierMother (see Introduce ensemble models by combining trained models from MotherCV #90)

• Combine fitted estimators from cross-validation runs into an ensemble classifier

• Provide default option for aggregation of member predictions: averaging without weighting

• Provide options for estimation of knowledge uncertainty:

  1. variance: computes variance of ensemble predictions over classes and members to represent ensemble uncertainty
  2. disagreement: computes disagreement score using KL-divergence to represent ensemble uncertainty
  3. member_average: computes ensemble uncertainty by averaging individual member uncertainties

• Fix estimator handling in mother_cv when tuning is disabled (see # mother_cv returned identical member estimators without nested tuning #92)

• Add example notebook demonstrating usage

• Add unit tests

Motivation

The goal is to enable ensemble predictions from models produced during cross-validation while keeping the existing workflow for model training and evaluation.

Testing

• Added automated tests covering ensemble classifier behavior
• Verified existing test suite continues to pass
• Validated example notebook

Notes

This is currently a draft PR.
I may still make minor adjustments and rebase onto the latest upstream main before requesting formal review.

Open To do's

• review unit tests covering 'mother_cv' usage for ranking with CatBoost - 2 tests failing right now

Follow-up tasks

• Provide more options for aggregation: weighted average, majority vote, etc.
• Provide more options for ensemble uncertainty estimation: BALD/ mutual information, etc.
• Add calibration methods for probability and uncertainty calibration of individual ensemble members based on their cv validation fold from fitting with mother_cv