autodoc: Execute .pyi stubs with deferred annotation semantics

[bfs2021]

Resolves #14354

What

When autodoc imports a PEP 561 .pyi stub file for a native module (via _find_type_stub_spec/_StubFileLoader), it now executes the stub as if it began with from __future__ import annotations — which is exactly how PEP 484 defines stub-file semantics ("The type checker is required to treat annotations in stub files as if they were always evaluated with from __future__ import annotations"). Previously the stub was compiled and executed as a plain module, so a forward reference in an annotation raised NameError at import time:

File ".../coconext/types.pyi", line 17, in <module>
    class Logic:
File ".../types.pyi", line 19, in Logic
    def __init__(self, arg: Logic) -> None: ...
NameError: name 'Logic' is not defined

The implementation compiles the stub source with the CO_FUTURE_ANNOTATIONS compiler flag instead of spec.loader.exec_module, so annotations stay lazily-evaluated strings and the module namespace is populated exactly as before otherwise.

Testing

• tests/roots/test-ext-apidoc-duplicates/fish_licence/halibut.pyi (previously empty) now contains a class with a self-referencing annotation, which reproduces the reported NameError on main.
• test_import_native_module_stubs_defer_annotations asserts the import succeeds and the annotation is preserved as a deferred string ('Fish').
• Verified the new test fails on main with the exact reported error (NameError: name 'Fish' is not defined) and passes with this change.

ruff check and ruff format --check (pinned 0.14.9) are clean. The remaining test_ext_autodoc failures in my local run (test_overload3, test_final, test_autodoc_pep695_type_alias, test_cfunction) are identical on unmodified main in this environment and unrelated to this change.

[bfs2021]

Heads-up on the red CI: the failing jobs here (Docutils HEAD, LaTeX, Python 3.13/3.14 — all failing on test_cfunction, where time.asctime's parameter renders as time_tuple instead of tuple) appear to be runner/dependency drift rather than anything from this change. The exact same job set fails on unrelated open PRs right now, e.g. #14726 (Docutils HEAD, LaTeX, Python 3.13t/3.14t) and #14724 (same).

The checks that exercise this change all pass: ruff, mypy, ty, Windows, macOS, docs-lint, and the ReadTheDocs build. The regression test also fails on main without this patch and passes with it (details in the description).