Fix #29929.

I implemented my suggestion in #29929 (comment) and added a simple unit test.

The unit test seems to fail on main consistently. Note that this relies on the values of np.empty so this is not fully deterministic but on my machine it seems like the following snippet consistently produces a warning:

import numpy as np

np.empty(1000).astype(np.int64)

A more realistic snippet that gives the warning is below. On my machine this produces a RuntimeWarning roughly 3-4 times out of 10 but takes a lot longer than the unit test (each run takes maybe ~2 seconds).

import numpy as np

from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import GridSearchCV
from sklearn.datasets import make_classification

X, y = make_classification(n_samples=10)
lr = LogisticRegression()
grid = GridSearchCV(lr, cv=2, param_grid={'max_iter': range(40, 240)})
grid.fit(X, y)