[python-package] Guard docstring assignments for python -OO compatibility

[PavelGuzenfeld]

Summary

Running XGBoost with python -OO (PYTHONOPTIMIZE=2) strips all docstrings,
making __doc__ attributes None. This causes an AttributeError crash at
import time:

File "xgboost/sklearn.py", line 1826
    fit.__doc__ = XGBModel.fit.__doc__.replace(
AttributeError: 'NoneType' object has no attribute 'replace'

This PR wraps each __doc__ assignment with a None check instead of removing
them, so documentation is fully preserved in normal mode and the crash is
avoided under -OO.

Why this approach (guard, not remove)

PR #12085 was correctly rejected —
it removed the assignments outright, which would have dropped method-level
docstrings for save_model, load_model, XGBClassifier.fit, and Dask wrappers
in normal Python runs.

This PR instead adds a one-line if __doc__ is not None: guard before each
assignment. The result:

Mode	Before this PR	After this PR
Normal python	Docstrings work	Docstrings identical, no change
python -O	Docstrings work	Docstrings identical, no change
python -OO	ImportError crash	No crash, no docstrings (expected)

Addressing maintainer concern from #8537

"adding checks to all those places, and having to run all the tests with this
additional option is infeasible" — @trivialfis

There are only 5 assignments across 2 files. Each fix is a single if guard.
No new tests needed — the fix is defensive by nature and cannot alter behavior in
normal mode (the guarded code is identical to the original).

Changes

• python-package/xgboost/sklearn.py — guard save_model.__doc__, load_model.__doc__, and XGBClassifier.fit.__doc__
• python-package/xgboost/dask/__init__.py — guard DaskXGBClassifier.predict_proba.__doc__ and DaskXGBRanker.fit.__doc__

Closes #8537

[RAMitchell]

Why are there 2 PRs on this?

[PavelGuzenfeld]

The two PRs address the same issue (#8537) but take different approaches:

• [python-package] Fix import crash under python -OO by removing outdated docstring assignments #12085 — removed the docstring assignments entirely. This was rejected because it would drop method-level docstrings (save_model, load_model, XGBClassifier.fit, Dask wrappers) in normal Python runs.
• [python-package] Guard docstring assignments for python -OO compatibility #12093 — guards each assignment with if __doc__ is not None:, preserving full docstring behavior in normal mode while preventing the crash under python -OO.

#12085 is closed and superseded by #12093.

[PavelGuzenfeld]

@RAMitchell Could you please reopen this PR? It was likely closed due to the confusion about the two PRs, but as clarified above, #12085 (the removed approach) is already closed and this one (#12093) is the corrected fix with a different approach.

The underlying issue (#8537) is still present — XGBoost crashes on import under python -OO. This PR adds minimal if __doc__ is not None: guards (5 lines across 2 files) to fix it.