Fix GapEncoder(init="k-means") crash on columns containing nulls

[Hrafz]

Addresses issue
#2237

Description

GapEncoder(init="k-means") crashes on any column containing a missing
value, while the two other init options fit the same column fine:

import pandas as pd
from skrub import GapEncoder

s = pd.Series(["alice", "bob", None, "alice", pd.NA, float("nan")], name="s")

GapEncoder(n_components=3, init="k-means++").fit_transform(s)  # OK
GapEncoder(n_components=3, init="random").fit_transform(s)     # OK
GapEncoder(n_components=3, init="k-means").fit_transform(s)
# AttributeError: 'NAType' object has no attribute 'lower'

Expected: init="k-means" handles the column's nulls the same way the other
two init options do (nulls are treated as the empty string throughout
GapEncoder, see skrub._utils.unique_strings and the existing
test_missing_values).

Actual: it raises from scikit-learn's text preprocessor. The exception type
depends on the column's dtype rather than on which sentinel appears first:

column	init="k-means" raises
object dtype containing np.nan	ValueError: np.nan is an invalid document, expected byte or unicode string.
object dtype containing None	ValueError: np.nan is an invalid document, expected byte or unicode string.
string dtype containing pd.NA	AttributeError: 'NAType' object has no attribute 'lower'

init="k-means++" and init="random" fit all three without error.

Cause: _init_vars builds the null-substituted array unq_X via
unique_strings(X, is_null) and uses it everywhere else in the method, but
was calling self._init_w(unq_V[lookup], X) with the raw X (still
containing the original null). Only the "k-means" branch touches X
directly — it re-vectorizes it from scratch inside get_kmeans_prototypes,
which is where the null reaches the preprocessor. "k-means++" and
"random" only ever consume the already-vectorized V.

Fix: pass unq_X[lookup] instead of X — the same null-substituted,
per-sample reconstruction already used throughout the rest of the method.
When there are no nulls, unq_X[lookup] == X, so this is a no-op for the
common case.

Checklist

• I have read the contributing guidelines
• I have added tests that verify the bug fix
• I have added an entry to CHANGES.rst describing the fix
• My code follows the code style of this project
• I have checked my code and corrected any misspellings

How Has This Been Tested?

test_missing_values is now parametrized over init, so it also exercises
"k-means" (previously only the default "k-means++").

Before the fix (test change only, source untouched):

skrub/tests/test_gap_encoder.py::test_missing_values[pandas-numpy-dtypes-k-means] FAILED
skrub/tests/test_gap_encoder.py::test_missing_values[pandas-nullable-dtypes-k-means] FAILED
2 failed, 4 passed, 44 deselected

After the fix:

skrub/tests/test_gap_encoder.py -q
50 passed

The full skrub/tests suite passes with no failures and no change in the
skipped/xfailed/xpassed counts beyond the newly-passing parametrizations.

[rcap107]

Hi @Hrafz, thanks for the PR. We'll be looking at this and #2240 in the next few days (most maintainers were off in the past few weeks)