`threading.local()` access should scale well from multiple threads

[colesbury]

Currently, accessing a threading.local() from multiple threads doesn't scale well because of reference count contention on the shared _thread._local object. We should use deferred reference counting on _thread._local to avoid this bottleneck.

Linked PRs

• gh-128691: Use deferred reference counting on _thread._local #128693
• Revert "gh-128691: Use deferred reference counting on _thread._local (#128693)" #128753