Decouple from dowhy internals & adopt new econml/dowhy features (follow-up to #338)

[ZmeiGorynych]

Follow-up to the econml 0.16 / dowhy 0.14 upgrade (#338). That PR was intentionally behavior-preserving; this issue tracks the deferred modernization and feature-adoption work.

1. Decouple from dowhy internals

The bump kept CausalTune's existing coupling to dowhy internals and only adapted it. Durable-decoupling items:

• Migrate EconML dispatch off string method_name. CausalTune dispatches via causal_model.estimate_effect(method_name="backdoor.econml.dml.LinearDML", …), which dowhy ≥0.13 deprecates (we currently just silence the warning in causaltune/__init__.py). Note CausalModel.estimate_effect unconditionally overwrites method_params["econml_estimator"] with the string, so the only instance-based path is to bypass it and call dowhy's functional estimate_effect(data, treatment, outcome, identifier_name, estimator=<Econml adapter instance>, …). Moderate effort + core-path test coverage; removes the deprecation for good.
• Reduce reliance on private attributes. shap.py / visualizer.py read econml's estimator._input_names["feature_names"]; scoring.py reads _effect_modifier_names / _treatment_value. Introduce a thin, documented adapter surface instead of reaching into privates.
• Reconsider the effect_tt base-class monkey-patch and effect_stderr. dowhy 0.14's Econml adapter now natively provides effect_tt / apply_multitreatment / effect_inference. Consider subclassing the adapter (or a dedicated estimator wrapper) rather than patching CausalEstimator.
• Custom-estimator plumbing. DoWhyWrapper still constructs CausalEstimate by hand and does not call super().__init__. Consider aligning it with dowhy 0.14's estimator lifecycle to reduce future breakage.
• Forward control_value to custom estimators' fit (Codex, deferred from Upgrade econml to 0.16.0 and dowhy to 0.14 (numpy 2, Python 3.10–3.12) #338). dowhy 0.14 passes control_value only to estimate_effect, so DoWhyWrapper.fit constructs the inner estimator with control_value=0. Harmless today (control is ~always 0), but should be threaded through for non-zero-control datasets.

2. Adopt new capabilities

Evaluate and, where they fit the AutoML loop, adopt:

• dowhy (0.10 → 0.14): doubly-robust estimator; refutation-based scoring / faster refuters as first-class model-selection signals; DRtester-style validation; generalized adjustment criterion for automatic adjustment-set selection; GCM missing-data support.
• econml (0.14 → 0.16): policy learning; new / updated CATE estimators; RScorer improvements; updated inference (effect_inference / BootstrapInference).

3. Broaden estimator coverage

• The experimental backdoor.propensity_score_weighting estimator is excluded by default and its effect_tt scoring path is not covered; decide whether to support or drop it.
• Revisit OrthoForest large-dataset handling now that dependencies are current.

4. Python 3.13

• Raise the ceiling to Python 3.13 once econml / numba / the rest of the stack publish 3.13-compatible wheels (the bump caps at <3.13).

Notes / gotchas discovered during the bump

• dcor ≥0.7 changed pairwise_distances(A, B) to return shape (len(B), len(A)) (transposed vs. psw_joint_weights); fixed in psw_energy_distance but worth an audit of any pairwise-distance usage added later.
• sklearn 1.6 strictly enforces feature-name order between fit/predict — the propensity scoring path had a fit/predict column-list mismatch (fixed in Upgrade econml to 0.16.0 and dowhy to 0.14 (numpy 2, Python 3.10–3.12) #338). Watch for the same pattern elsewhere.
• numpy 2 removed numpy.distutils and np.trapz; the stack now caps pandas <3 and scikit-learn <1.7, driven by FLAML / numba / econml ceilings.
• Several dataset loaders fetch from third-party URLs that are periodically dead (NHEFS host now serves HTML); test_datasets skips on download failure.

Basis: docs/followup_decouple_and_features.md (kept local in the PR branch).