fix(monkeypatch): restore attributes of objects with a custom __setattr__ - #15118

closes #15099

Problem

MonkeyPatch.setattr() records the old value from the instance __dict__ whenever no data descriptor is in the way, on the assumption that setattr() writes there:

elif not _is_data_descriptor(type(target), name):
    target_dict = getattr(target, "__dict__", None)
    if isinstance(target_dict, Mapping):
        oldval = target_dict.get(name, NOTSET)

A class which defines __setattr__ can store the value anywhere -- a proxy object's private mapping, for example. For those, oldval is recorded as NOTSET, so undo() calls delattr() instead of restoring the value. That raises AttributeError during teardown, and the patched value then leaks into later tests:

ERROR test_patch - AttributeError: 'Config' object has no attribute 'debug'
FAILED test_restored - assert True is False

Fix

Restrict the __dict__ lookup to types which use object.__setattr__, the only case where setattr() is known to write into the instance __dict__. For every other type the getattr() value stays the right one to restore, because undo() writes it back through the same __setattr__ hook.

This keeps what #14969 added for #10644 intact: those cases use plain classes, so they still take the __dict__ path and undo() still removes the entry it created.

Tests

• testing/test_monkeypatch.py::test_undo_attribute_stored_outside_instance_dict, a new regression test modelled on the reproducer in the issue. Without the source change it fails with AttributeError; with it, undo() restores the original value.
• changelog/15099.bugfix.rst, the news fragment.

Executed before submitting:

• pytest testing/test_monkeypatch.py -> 46 passed, 1 skipped
• pytest testing/ -> 4632 passed, 51 skipped, 15 xfailed, 5 xpassed, 1 failed, 1 error. Both the failure (test_fixture_doctest_skip_has_line_number) and the error (test_cache_makedir) reproduce identically on the unmodified checkout in this environment, so they are unrelated to this change.
• ruff check and ruff format --check with 0.16.9, the version pinned in .pre-commit-config.yaml: clean.
• mypy: no new diagnostics for the changed files. (The one remaining report, an unused type: ignore further down testing/test_monkeypatch.py, is present on the unmodified file too.)

AI assistance disclosure

Per the project's AI/LLM-Assisted Contributions Policy: this change -- the reproduction analysis, the fix, the regression test and the changelog entry -- was produced with AI assistance, and I validated it by hand before submitting. I reproduced the failure and confirmed the fix on Python 3.13, verified the new test fails without the source change, and ran the full suite plus ruff and mypy. I am happy to go through any part of the diff.