Migrate OrdinalEncoder.fit() to narwhals, add polars support

[solegalli]

Depends on #1047 (shared encoder helper and init checks).

Migrates OrdinalEncoder.fit() to narwhals with polars support. transform() / inverse_transform() already come dataframe-agnostic from CategoricalMethodsMixin.

fit() has two paths: "arbitrary" (X[var].unique()) and "ordered" (target mean per category via y.groupby(X[var])).

Merge vs split: benchmarked a pure-narwhals fit() at 10k–100k rows × 1–10 cols × 5–50 categories — 5x–18x slower than pandas-native at every size (a large, consistent loss, unlike the ~1.1x for the transform hot path). So fit() splits on is_pandas = nwd.is_pandas_dataframe(X): pandas keeps a close variant of its groupby/unique code (like-for-like benchmark within noise of the old code), polars/other backends go through group_by()/agg()/sort()/unique(). "arbitrary" is untouched.

Two issues, both confirmed pre-existing against the unmodified file:

1. Bug (fixed): "ordered" always called y.groupby(X[var]), raising AttributeError on a numpy y (list/array-like target, as sklearn's check_X_y produces). This is exactly what test_encoders_when_x_pandas_y_numpy exercises for OrdinalEncoder (encoder2, added 2022 for Encoders that are f(X, y) can produce nan results when y has non-standard index and X becomes an np.ndarray #376) — failed on the unmodified file, now passes. Fixed on both branches (pandas: .assign() to pair X[var] with y; narwhals: nw.new_series).
2. Cross-backend ordering hazard (avoided): grouping by category then sorting by target mean doesn't guarantee the same tie-break across backends (polars reversed two tied categories vs pandas). Old pandas code effectively tie-broke on the category itself; reproduced explicitly with a compound .sort([target_name, var]). Verified pandas and polars now produce the same dict for a tied-mean fixture, matching the old order.

Tests: every test rewritten as one make_df in [pd.DataFrame, pl.DataFrame] case per behaviour (43, up from 26). test_variables_cast_as_category stays pandas-only.

Verified: test_ordinal_encoder.py 43 passed; full tests/test_encoding 344 passed / 16 failed — identical IDs to the unmodified base (17; one being the bug fixed here). flake8 / mypy clean, sphinx -W clean. OrdinalEncoder.rst examples verified against real output (California Housing), "With polars" section added; Titanic examples untouched (no network in sandbox).

[solegalli]

Updated this branch:

• Rebased onto Add shared backend test fixtures and helpers for narwhals-migrated tests #1045 (shared backend test fixtures and helpers). This PR is now stacked on Add shared backend test fixtures and helpers for narwhals-migrated tests #1045, so its diff includes that PR's commits until it merges.
• Tests rewritten to the shared structure: make_df and the data_enc* fixtures from tests/test_encoding/conftest.py, isinstance(X, make_df) plus to_dict() checks, and pytest.raises/warns(match=re.escape(msg)).
• The target is now built with make_series (pandas/polars Series), plus a test passing it as a list and as a numpy array.
• No source changes.

Locally: test_ordinal_encoder.py 47 passed; tests/test_encoding has no failures beyond those already on narwhals-migration.