Preserve unread lazy class annotations on Python 3.14

[pralav-25]

Unread annotations on a Python 3.14 dynamic class are lost when the class is pickled by value. A fresh-process round trip of class C: x: int returns an empty annotation dictionary unless its annotations were accessed first.

Materialize the native annotation cache before sorted class-state extraction, only for an owned lazy annotation function whose cache has not been populated. The snapshot retains the existing cache representation while excluding the annotation closure. Fresh-process regressions cover unread/read annotations, class-local/self/resolved references, derived classes, cached metaclass getters/setters and deterministic chained pickling.

Annotations are evaluated when dumping, consistent with dynamic function annotations. Unresolved bare names therefore raise NameError during dumping; quoted references and future-import annotations retain their string values. This behavior is recorded in the changelog.

Fixes #608.

Validation:

• Exact upstream source with the final regressions: 10 expected failures and 12 passes (eight annotation-loss failures and two checks of the documented NameError change).
• Python 3.14.8: 285 passed, 13 skipped and two expected xfails.
• Tox Python 3.12.14: 270 passed, 28 skipped and two expected xfails.
• Pinned pre-commit hooks, wheel/sdist build, Twine and git diff --check pass.
• Separate review passed 72 semantic scenarios and 64 native-cache chained cases across Python 3.14.8 and 3.15.0rc2 free-threaded. The full supported Python/OS matrix was not run locally.
• Four exploratory future-import byte-determinism cases reproduce on unchanged upstream; their annotation values are preserved. These baseline cases are separate from the native-cache path changed here.

AI assistance: OpenAI Codex prepared the implementation, regression tests and changelog, ran the checks, and performed a separate agent review of the final patch.