fix: Bypass deprecated feature_dtypes_ setter in fit()

[Jan Tilly (jtilly)]

Two changes:

• Instead of catching the warning, I'm assigning directly to _feature_dtypes_ instead of going through the deprecated feature_dtypes_ setter, eliminating the DeprecationWarning emitted on every fit() call when X is a DataFrame.
• I'm removing the deprecation warning from the feature_dtypes_ getter, because that's also firing for things such as dir(...) and other introspection calls.

import warnings
import pandas as pd
from glum import GeneralizedLinearRegressor
warnings.filterwarnings("error", message="Use `categorical_levels_` instead.")

X = pd.DataFrame({"a": [1.0, 2.0, 3.0], "b": [4.0, 5.0, 6.0]})
y = [1.0, 2.0, 3.0]
model = GeneralizedLinearRegressor().fit(X, y)

dir(model)

/tmp via py v3.12.13 ❯ pixi exec --spec glum python reproducer.py
Traceback (most recent call last):
  File "/private/tmp/reproducer.py", line 10, in <module>
    dir(model)
    ~~~^^^^^^^
  File "/tmp/rattler/cache/cached-envs-v0/glum-12c16ede7e8029a1/lib/python3.14/site-packages/sklearn/base.py", line 214, in __dir__
    return [attr for attr in super().__dir__() if hasattr(self, attr)]
                                                  ~~~~~~~^^^^^^^^^^^^
  File "/tmp/rattler/cache/cached-envs-v0/glum-12c16ede7e8029a1/lib/python3.14/_py_warnings.py", line 799, in wrapper
    _wm.warn(msg, category=category, stacklevel=stacklevel + 1)
    ~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
DeprecationWarning: Use `categorical_levels_` instead.