GH-16779: update estimators to support sklearn 1.6+

[zazulam]

Fixes #16779

Summary

h2o.sklearn wrappers need compatibility updates for newer scikit-learn APIs, especially around estimator type semantics and tags behavior in scikit-learn 1.6+. scikit-learn revamped estimator tags in 1.6.0 (December 2024) and introduced __sklearn_tags__ as the preferred API (with Tags objects). This affects third-party estimator wrappers and type-dispatch behavior (is_classifier, is_regressor, clone/check tooling).

For generic H2O sklearn wrappers (e.g. H2OGradientBoostingEstimator(estimator_type='classifier')), semantics can drift in sklearn integration paths (notably clone + tags/type checks) unless wrapper params and _estimator_type are propagated consistently.

Proposed / implemented fix

In h2o-py/h2o/sklearn/wrapper.py:

• Keep explicit _estimator_type precedence in classifier/regressor checks.
• Include estimator_type and init_connection_args in sklearn parameter flow (get_params/set_params) so clone semantics are preserved.
• Keep these params reserved from forwarding to underlying H2O estimators.
• Implement __sklearn_tags__ handling for newer sklearn tags API when available.

In h2o-py/tests/testdir_sklearn/pyunit_sklearn_api.py:

• Add regression tests for:
  • classifier/regressor identification
  • clone preserving estimator_type semantics
  • sklearn tags estimator_type behavior (when available)

Validation run

• testdir_sklearn/pyunit_sklearn_api.py → PASS
• testdir_sklearn/pyunit_sklearn_params.py → PASS

Dependency notes

• No H2O package dependency changes are required for this fix itself.
• Compatibility validation was performed with latest scikit-learn 1.x available in test env: 1.6.1.
• Local environment note: running older sklearn wheels with incompatible NumPy can cause ABI errors; this was an environment concern, not a required project dependency change.