gh-132261: Store annotations at hidden internal keys in the class dict

[JelleZijlstra]

This fixes #132261 and provides a more elegant solution for https://peps.python.org/pep-0749/#annotations-and-metaclasses. The trouble is that we store things at the __annotate__ and __annotations__ keys in the class dictionary and they interfere with the descriptors on type: the solution is to not store our internal objects at those keys. Instead, we now store the annotate function at __annotate_func__ and the annotations dict (once evaluated) at __annotations_cache__. These names remain undocumented implementation details.

Downsides:

• If you look at the class dict, you'll see those new keys. typing.py needed some updates to deal with this.
• The compatibility store with 3.13 changes in regards to looking directly in the class namespace for __annotations__. In 3.13, this always worked: if a class has annotations there's an __annotations__ key in the class dict. On current main, it sometimes works (only if somebody has accessed the .__annotations__ attribute). With this PR, it never works (unless the __annotations__ key was explicitly added). Breaking this pattern consistently seems better than breaking it nondeterministically though.
• It becomes harder to get the annotate function in a metaclass __new__. I made annotationlib.get_annotate_function() handle this.
• The internal names become accessible as attributes (e.g. you can do cls.__annotate_func__).

• Issue: [3.14] annotationlib - calling get_annotations on instances gets an unexpected error #132261

---

📚 Documentation preview 📚: https://cpython-previews--132345.org.readthedocs.build/

[carljm]

This looks pretty good to me! I expected it to be a lot gnarlier. It really just removes the ugly special handling of class annotations, everything it adds looks predictable and reasonably straightforward.

Thanks for your work on this!