Support pickling cached_property descriptors on Python 3.8-3.11

[charan-rathore]

Fixes #438.

On Python before 3.12, functools.cached_property stores an RLock. Pickling a dynamically defined class or instance with one of these properties fails because the lock cannot be pickled.

Add a reducer that reconstructs the descriptor from its function and restores its attributes, leaving the constructor to create a new lock. This follows pierreglaser's suggestion in the issue to add a minimal reducer. Cached results belong to the instance dictionary and continue to travel with the instance.

The reducer keeps the assigned attribute name, documentation and custom attributes. On Python 3.12 and newer, cached_property no longer creates a lock, so a custom attribute named lock is preserved rather than removed.

Tests cover classes in a fresh process, instances before and after computing a cached result, standalone descriptors, repeated references and descriptors not yet assigned to a class. Both default and protocol-2 test classes run them.

Verification: Python 3.10 and 3.11 - the 8 new tests fail on master and the patched full suite passes (3.10: 281 passed, 3.11: 276 passed). On Python 3.12 to 3.14 the new tests already pass on master and the patched suites pass. Python 3.8 and 3.9 were not tested. These are same-version roundtrips; cloudpickle expects the communicating processes to run the same Python version. Only the exact functools.cached_property is handled, not subclasses.