Migrate ProbeFeatureSelection to narwhals, add polars support

[solegalli]

Stacked on #1070 (selection base classes). Its commit shows in the diff until #1070 is merged.

Summary

ProbeFeatureSelection now accepts pandas, polars and other dataframes supported by narwhals, and returns the same library it receives.

• fit(): nw_X, y = check_X_y(X, y); the native X goes to _select_numerical_variables and _get_feature_names_in.
• _generate_probe_features(n_obs) returns a dict {name: numpy array}. The numpy calls and their order did not change, so the probe values for a given random_state are the same as before.
• Adding the probes:
  • pandas: DataFrame(probes, copy=False) + concat([X[variables_].reset_index(drop=True), probes], axis=1), called through nw.get_native_namespace(X) like _importance_series does, so there is no pandas import.
  • other backends: nw.from_dict(probes, backend=...) + nw.concat(..., how="horizontal").
• fit() no longer runs X.reset_index(drop=True, inplace=True). It was safe only while check_X copied the input. Now that check_X does not copy, it would reset the user's index.
• collective=False builds feature_importances_ / feature_importances_std_ with _importance_series, as the base does. They are a pandas Series for pandas input and a dict for other input.
• _get_features_to_drop(probes) takes the probe names and reads the importances by key, so it works with both the Series and the dict. For a single probe it still compares against the 1-element array, which keeps the old behaviour for the degenerate cases described below.
• probe_features_ is a dataframe in the input's library.
• The docstring had the distribution option as 'binomial', which raises an error. It now says 'binary', the valid name.

Benchmarks

The machine was heavily loaded (load average about 45 on 10 cores), so the numbers are noisy. Each value is the median of 21-25 alternating repeats, in ms. Cross-validation takes almost all of fit(), so I timed the part the selector does itself: building the probe frame and appending it to X, with the probe arrays already generated.

pandas: DataFrame(copy=False) + concat is the fastest at every size.

rows	cols	probe cols	old (df[c]=arr loop + concat)	DataFrame + concat	DataFrame(copy=False) + concat	assign	narwhals concat	narwhals with_columns
10k	10	1	0.43	0.38	0.20	0.23	1.56	2.89
10k	50	12	2.09	0.57	0.24	2.07	5.05	12.75
10k	200	12	2.42	1.21	0.35	2.01	35.5	40.8
100k	10	12	3.44	3.23	0.52	5.40	5.48	7.39
100k	50	1	1.84	0.58	0.30	0.64	9.23	14.84
100k	200	12	2.73	3.23	0.41	4.00	40.3	80.5
500k	10	12	9.86	11.42	0.39	18.46	41.2	30.9
500k	50	12	9.32	13.21	0.36	18.87	46.8	46.6
500k	200	12	10.42	11.19	0.33	16.94	42.5	61.2

The two narwhals columns come from a separate run, so compare them with the other columns only roughly.

fit() with the cross-validation replaced by a stub (the selector's own work). Old = origin/main, whose check_X still copied X. Each cell lists 3 runs, each the median of 9 fits, in ms:

rows	cols	probes	old	new
100k	50	normal x1	8.4 / 28.2 / 11.7	3.6 / 1.7 / 3.0
100k	50	all x2	41.8 / 25.5 / 51.1	26.2 / 28.8 / 27.2
500k	10	normal x1	20.7 / 26.4 / 17.7	10.2 / 49.2 / 10.9
500k	50	normal x1	47.6 / 45.8 / 117.9	12.4 / 12.1 / 11.9
500k	50	all x2	184.9 / 219.8 / 121.3	126.7 / 100.3 / 126.9

polars: appending the probes costs well under 1 ms at 500k rows and more. Generating the probes with numpy costs 10-1000 ms, and polars-native hstack and narwhals concat are within noise of each other:

rows	cols	probe cols	numpy generation	polars hstack	narwhals concat	narwhals with_columns
500k	10	1	10.2	0.06	0.10	0.14
500k	50	12	253	0.37	0.63	0.95
500k	200	12	265	0.69	0.60	0.77
2M	10	12	1062	0.37	0.36	0.47
2M	50	12	1016	0.33	0.31	1.13
2M	200	1	72	5.25	2.26	3.50

I kept narwhals concat for the non-pandas path. It also covers the other backends, while hstack would need a third, polars-only branch to save at most about 0.3 ms.

Behaviour

• pandas: identical to origin/main in 112 recorded cases. The cases cover:

  • collective True/False, every distribution, 'all' and lists, n_probes 1 and 3, the 3 thresholds
  • unseeded estimators, variables, confirm_variables, a splits generator, groups
  • NaN with HistGradientBoosting, a mixed numerical/categorical frame
  • integer column names, a non-default index, distribution=[] and n_probes=0

  In each case these match exactly: probe_features_ (values, dtypes, RangeIndex), feature_importances_ and _std_ (values, index), features_to_drop_, variables_, the transform output, the error messages, and the user's X staying untouched (same index).

• polars: same values as pandas, bit for bit, in the 108 applicable cases. probe_features_ is a polars DataFrame, and the importances are dicts (see below).

• All outputs in the user guide and the docstring were re-run: unchanged. I added the missing dtype: float64 line to one output.

Tests

tests/test_selection/test_probe_feature_selection.py is rewritten to the conventions:

• init: one error test per message, with wrong values and wrong types, using match=re.escape, plus test_init_param_assignment.
• fit/transform on both backends via make_df, using the data_classification fixture that is already in conftest.py. The expected importances are kept from the old tests.
• pandas-only tests for integer column names and for leaving a non-default index alone (that one fails if fit resets the index in place).
• probe generation and _get_features_to_drop unit tests, with both the Series and the dict container.

tests/test_selection failing tests: 137 on origin/narwhals-selection-base, 128 on this branch, no new failures. The 9 tests that now pass are the 5 old probe tests and 4 test_check_estimator_selectors cases for this class. tests/parametrize_with_checks_selection_v16.py -k Probe: the same 21 failures before and after. They are caused by numpy input to check_X, which fails for every migrated transformer.

flake8 feature_engine tests is clean. mypy feature_engine shows the same 2 errors as the base (in transformation/log.py).

Needs decision

• For polars input, feature_importances_ / feature_importances_std_ are dicts. This follows the base container from Migrate the selection base classes and helpers to narwhals, add polars support #1070 and its open question.

Pre-existing issues, not fixed

• fit() calls np.random.seed(random_state) and so resets numpy's global random generator. This makes unseeded estimators such as RandomForestClassifier() deterministic, and the user guide and tests rely on it. Switching to np.random.RandomState(random_state) would give the same probe values, but results with unseeded estimators would change, so I left it.
• distribution=[] or n_probes <= 0 pass __init__, and then fit() fails with numpy's "The truth value of an empty array is ambiguous". The fix would be to validate both in __init__, which changes when and how the error is raised.
• pandas input with integer column names and collective=True fails in scikit-learn, because the frame then mixes int and str column names (the probes). collective=False works.
• n_probes=True is accepted as 1, because bool is a subclass of int.