Fix linear selection on Python 3.12

[russlan23]

Fixes #24.

scikit-learn removed the multi_class parameter from LogisticRegression. This keeps the existing liblinear one-vs-rest behavior by wrapping the two logistic estimators in OneVsRestClassifier, then preserves the existing coefficient aggregation when computing feature scores.

Validation:

• Python 3.12.6 / scikit-learn 1.9.0: none, lasso, and ridge classification paths fit and transform successfully on Iris
• Python 3.11.10 / scikit-learn 1.5.2: the new unregularized scores match the legacy implementation exactly
• Python compile check and git diff check pass

[skadio]

@russlan23 thank you so much for this PR and the fix. Coincidently, we were also just in the middle of updating this specific issue. Please stay tuned for an update shortly. We might use this PR (or, close this one and make an version upgrade internally). Either way, this is going to be fixed very soon. cc: @rbaral @takojunior

[rbaral]

@russlan23 thanks for the suggestion you provided. This and with some other minimal fix in text_based.py (for upgraded pandas due to Python 3.12), I was able to run the test cases for Python 3.12. and Python 3.10. I can bundle these two fixes and some housekeeping work into a new PR. We also need to add some documentation about the fix.

It would have been better to have minimal code changes, such as using a different solver. However, that required some changes in the test cases. CC @skadio @takojunior