Adding a TableVectorizer specialization for HistGradientBoosting

[jeromedockes]

Problem Description

by default the TableVectorizer one-hot encodes low-cardinality categories. In the common case that a scikit-learn HistGradientBoosting{Regressor,Classifier} is used as the downstream estimator, that is suboptimal.
The best thing to do is to rely on the gradient boosting estimator's built-in handling of categories.

In scikit-learn 1.3 we need to transform them with an OrdinalEncoder and let the estimator know which features are categorical with categorical_features=["column_x", "column_y"]

In scikit-learn 1.4 we need to make sure those features have a categorical dtype in the transformer's output but otherwise leave them as they are, and the estimator will recognize and encode them appropriately if it is initialized with categorical_features="from_dtype"

Feature Description

maybe a subclass of TableVectorizer that has different default parameters.
In a first step we can address the scikit-learn >= 1.4 case which is the easiest.
In older scikit-learn versions the user needs to set categorical_features on the gradient boosting estimator, which is not something the tablevectorizer can do

Alternative Solutions

No response

Additional Context

No response

[GaelVaroquaux]

Excellent! I would also suggest to use the MinHashEncoder instead of the GapEncoder for the high-cardinality strings. It is faster and leads to better predictions with tree-based models.

[TheooJ]

Implemented in #926