import_importlib leaks module attr sets

Bug report

Bug description:

Our CI (https://github.com/facebookincubator/cinder/actions/runs/29382592631/job/87249270535) was failing on test_pickle and test_pickletool with:

PicklingError: Can't pickle
<class 'importlib._bootstrap.BuiltinImporter'>:
it's not the same object as
importlib._bootstrap.BuiltinImporter

I believe this is happening because we run multiple test modules in the same interpreter, so the following sequence happens:

• test_importlib.util.import_importlib() imports a source copy of importlib while blocking _frozen_importlib. During this import, importlib._bootstrap._setup() initializes import metadata on existing built-in modules. If builtins.__loader__ or builtins.__spec__ was originally absent, the source copy installs its own BuiltinImporter.
• Although import_fresh_module() restores sys.modules, it does not restore attributes mutated on existing module objects. Consequently, builtins.__loader__ continues to reference the temporary source BuiltinImporter, while importlib._bootstrap.BuiltinImporter resolves to the restored frozen class.
• Pickle serializes classes by module and qualified name and verifies that the resolved global is the same object. The two BuiltinImporter class objects therefore cause the identity check to fail.

To fix this we need to snapshot __loader__ and __spec__ before importing importlib and then restore them to their original values after.

CPython versions tested on:

CPython main branch

Operating systems tested on:

macOS

Linked PRs

• gh-154276

Start with test_importlib.util.import_importlib(), then reproduce the failure in test_pickle and test_pickletool. Trace how the temporary importlib import changes builtins.loader and builtins.spec; done means the original values are restored and both tests pass without the PicklingError. A linked PR, gh-154276, indicates work is already underway.