fix: surface MultiOutputRegressor guidance when treatment_featurizer + single-output model_t (closes #1012)

[immu4989]

Closes #1012.

What's happening

When treatment_featurizer produces a multi-column featurized treatment (e.g. PolynomialFeatures yielding [T, T**2, ...]), the first-stage model_t receives a 2D Y target. If the user-supplied regressor doesn't support multi-output (CatBoost, older XGBoost pre-1.6, etc.), the underlying estimator raises an opaque shape error from inside _fit_with_groups, far from the actual cause.

The reporter asks: "Is this the intended behavior? Can a model be fitted to each polynomial degree of the feature or do I need a multi-output regressor?"

Answer is the latter, but they shouldn't have had to ask.

Fix

_FirstStageSelector.train (econml/dml/dml.py) now catches ValueError from the underlying self._model.train(...) call. When the target is multi-column (ndim == 2 and shape[1] > 1) and is not a discrete-treatment one-hot encoding, the error is re-raised with:

First-stage model failed to fit a N-column target. This typically happens when treatment_featurizer (or a multi-dimensional outcome) produces a multi-column target but the supplied model does not support multi-output regression. Wrap your model with sklearn.multioutput.MultiOutputRegressor, or use a model with native multi-output support (e.g. LinearRegression, RandomForestRegressor, GradientBoostingRegressor). Original error: <underlying message>

The original error is preserved via raise ... from exc so debugging is unaffected.

Verification

New test_single_output_model_t_with_featurizer_raises_helpful_error (in test_treatment_featurization.py):

• Builds a mock _SingleOutputOnlyRegressor that errors on 2D y (mirrors CatBoost's failure mode)
• Asserts the raised ValueError mentions MultiOutputRegressor and the multi-output framing
• Then re-runs the same setup with MultiOutputRegressor(_SingleOutputOnlyRegressor()) and asserts fit succeeds and effect() returns finite values

Pre-fix: test fails with opaque expected 1D target, got shape (300, 2).
Post-fix: test passes.

Broader sanity: 28 tests passed across test_dml.py + test_treatment_featurization.py (ray-marked tests deselected per the known env issue on main).