- update to 1.3.2:
* All dataset fetchers now accept `data_home` as any object that
implements the :class:`os.PathLike` interface, for instance,
:class:`pathlib.Path`.
* Fixes a bug in :class:`decomposition.KernelPCA` by forcing the
output of the internal :class:`preprocessing.KernelCenterer` to
be a default array. When the arpack solver is used, it expects
an array with a `dtype` attribute.
* Fixes a bug for metrics using `zero_division=np.nan`
(e.g. :func:`~metrics.precision_score`) within a paralell loop
(e.g. :func:`~model_selection.cross_val_score`) where the
singleton for `np.nan` will be different in the sub-processes.
* Do not leak data via non-initialized memory in decision tree
pickle files and make the generation of those files
deterministic.
* Ridge models with `solver='sparse_cg'` may have slightly
different results with scipy>=1.12, because of an underlying
change in the scipy solver
* The `set_output` API correctly works with list input.
* :class:`calibration.CalibratedClassifierCV` can now handle
models that produce large prediction scores.

- Skip another recalcitrant test on 32 bit.
* We are in the process of introducing a new way to route metadata
such as sample_weight throughout the codebase, which would
affect how meta-estimators such as pipeline.Pipeline and
* Originally hosted in the scikit-learn-contrib repository,
* A new category encoding strategy preprocessing.TargetEncoder
encodes the categories based on a shrunk estimate of the average
* The classes tree.DecisionTreeClassifier and tree.DecisionTreeRegressor