Encode numeric categorical columns with a numpy ordinal encoder

[Innixma]

Stacked on #1255.

EncodeCategoricalFeaturesStep's ordinal path wrapped sklearn's OrdinalEncoder. On a small table almost all of a call is sklearn's input validation rather than the encoding, and the step fits once per ensemble member, so on a categorical-heavy table of a few hundred rows the step was the largest single CPU cost of fit. Profiled on a 598 x 31 table with 25 categorical columns and 88 ensemble preprocessing pipelines: 0.54 s of a 0.85 s fit, about 6 ms per pipeline, of which the encoder's own arithmetic is well under a millisecond.

Numeric arrays now go through NumericOrdinalEncoder, a OneToOneFeatureMixin transformer that does one np.unique and one np.searchsorted per column and validates nothing. It gives the same output as OrdinalEncoder(handle_unknown="use_encoded_value", unknown_value=np.nan): sorted distinct values become 0..k-1 as float64, NaN stays NaN, a value unseen at fit becomes NaN, and categories_ lists NaN last for a column that had NaN at fit so the per-column category count that the ordinal_shuffled permutation reads is unchanged. Object arrays keep sklearn's encoder. The ColumnTransformer wrapper, output names, schema carry-over and the one-hot path are untouched.

On the table above the step goes from 3.97 ms to 1.20 ms per fit_transform and from 1.59 ms to 0.27 ms per transform; preprocessed outputs are identical. Tests compare the new encoder with sklearn's on plain, missing-value, unknown-at-transform, float32, all-NaN and constant columns, check the encoder choice by dtype, and check the whole step's output, including the shuffled variant, against the step with sklearn's encoder inside. The four test_sklearn_estimator_checks failures for KDITransformerWithNaN in this environment are present on the base branch as well.

🤖 Generated with Claude Code

https://claude.ai/code/session_01ELdutHiUqkvynzEP7EnPsi

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 9be1cba. Configure here.}