Preserve explicit subclass overrides when pickling dynamic classes

[soodoku]

When a subclass explicitly assigns the same object as a base attribute, cloudpickle removes that assignment from the serialized class state. In a fresh process, updating the base then changes the subclass too, despite its explicit override. Fixes #584.

Preserve the class's own dictionary entries. Python already excludes merely inherited attributes from this dictionary; sharing an object with a base attribute does not make an explicit assignment redundant. The subprocess tests cover constants and methods, single and multiple inheritance, and protocols 2 and 5. They start a worker before defining the classes so class tracking cannot mask the failure.

Class construction already sets __module__, so state restoration skips a redundant assignment of that attribute. This avoids triggering custom metaclass setters twice while still restoring a module value that changed on an existing tracked class. Both cases have regression tests.

Reproducer (run as a script):

import subprocess
import sys
import cloudpickle

class Parent:
    version = 1

class Child(Parent):
    version = 1

payload = cloudpickle.dumps((Parent, Child))
Parent.version = 2
assert Child.version == 1

output = subprocess.check_output(
    [sys.executable, "-c", """
import sys
import cloudpickle
Parent, Child = cloudpickle.loads(sys.stdin.buffer.read())
Parent.version = 2
print(Child.version)
"""],
    input=payload,
)
assert output == b"1\n", output

On upstream 4af5936, the final assertion fails with b"2\n". It passes with this fix.

Validation on macOS:

• Python 3.12.11: baseline 266 passed, 10 skipped, 2 xfailed; candidate 274 passed, 10 skipped, 2 xfailed.
• Python 3.14.7: baseline 264 passed, 12 skipped, 2 xfailed; candidate 272 passed, 12 skipped, 2 xfailed.
• Reverting only the override-preservation change makes its four regression cases fail; removing only the module-assignment guard makes both metaclass cases fail. The module-restoration tests also reject an unconditional omission of __module__. All eight new cases pass with the final fix.
• Independent review by a different model in an isolated checkout found no remaining correctness issues after checking the metaclass and class-tracker cases. It independently ran the full suites and confirmed the regression failures with the fixes removed.
• All configured pre-commit hooks pass on the changed files. Ruff 0.14.3 passes across cloudpickle and tests.
• python -m build builds the source distribution and a wheel from it; twine check --strict dist/* passes. A fresh environment using the installed wheel, outside the source checkout, passes 274 tests, with 10 skipped and 2 xfailed. The README's functional examples and the reproducer pass against that wheel.

Full suites use PYTHONPATH='.:tests' python -m pytest -q with the development dependencies plus NumPy and SciPy. The wheel suite uses a separate copy of the tests because the existing source distribution omits them (#555).