Migrate TargetMeanRegressor to narwhals, add polars support

[solegalli]

Stacked on #1066 (BaseTargetMeanEstimator migration), which is not merged yet: until it is, this diff also shows #1066's changes. The only commit of this PR is the last one.

Summary

• TargetMeanRegressor no longer imports pandas. fit and predict are typed for any narwhals-supported dataframe, and fit takes the target as a series, numpy array or list. All the computation lives in the base class (Migrate BaseTargetMeanEstimator to narwhals, add polars support #1066); the regressor only adds the binary-target check (type_of_target) and predict/score.
• Docstring: mention of pandas and polars support, a polars example (run, real output pasted), a typo fix ("should of" -> "should be of").
• There is no user guide page for this class (the _prediction module is private), so no docs page changed.
• Tests rewritten to the conventions (make_df, make_series, full-message match=, explicit expected values).

Benchmarks

The regressor's own work is type_of_target(y). I compared it on the native target with type_of_target(np.asarray(y)), and against the whole fit (median of 7 / 5 runs, ms, float target, half numerical and half categorical columns):

backend	rows	cols	type_of_target(native)	type_of_target(numpy)	fit	predict	score
pandas	10k	4	0.3	0.1	10.5	6.5	2.1
pandas	100k	20	0.2	0.1	138.3	48.8	47.7
pandas	500k	4	0.6	0.5	129.6	46.8	48.5
pandas	500k	20	0.7	0.6	689.2	214.2	203.5
polars	500k	4	1.5	0.6	49.2	25.7	28.2
polars	1M	20	1.0	1.0	382.0	212.8	220.3
polars	2M	4	3.4	3.2	144.5	82.4	98.1
polars	2M	20	2.3	2.1	714.4	475.3	557.8

The check is under 2% of fit on both backends and converting to numpy first saves at most ~1 ms, so I kept type_of_target on the target as passed (no branch needed).

Behaviour

• pandas: identical to Migrate BaseTargetMeanEstimator to narwhals, add polars support #1066's head. I recorded predictions, score, encoder_dict_, binner_dict_ and error messages over 160 cases (both strategies, bins 1/3/10, 4 variable selections, series/list/numpy targets, binary target, NaN target, short target, unseen category, NaN in X, not fitted, reordered columns, integer column names, non-dataframe input) before and after: all equal.
• polars: same values as pandas in every one of those cases, with polars, numpy and list targets.

Tests

• tests/test_prediction: 122 passed before, 160 passed after, 0 failures.
• tests/test_selection (imports TargetMeanRegressor): 177 failed / 251 passed before and after, same failing tests (the selection module is not migrated yet).
• flake8 feature_engine tests clean; mypy feature_engine has the same 2 errors as the base ref (datetime_subtraction.py, log.py).
• The old test_classifier_numerical_variables checked num_var_B with np.array_equal(...) without assert; the new test asserts it.

Needs decision

• Mixing backends between X and y fails in fit, in add_target_to_X (shared helper, used by the base class): pandas X with a polars y raises AttributeError: PolarsSeries has not attribute '_broadcast', and polars X with a pandas y raises a narwhals ComputeError about a duplicate 'literal' column. Lists and numpy arrays work with both. Should mixed backends be supported (convert y to numpy when its backend differs from X's), or rejected with a clear error? I did not change the helper or the base here.
• The df_regression fixture in tests/test_prediction/conftest.py is no longer used (the data now lives in the regressor's test file, as it is used by one file only). I left it to avoid a conflict with the parallel TargetMeanClassifier PR, which edits the same fixture file; it can be removed after that one merges.