Fix free-threading crash from untraversed deferred references in suspended greenlets

[kumaraditya303]

Hi — I'm a CPython core developer working on free-threading and maintainer of asyncio.

This PR fixes the crash on free-threading caused by deference reference counted objects not traversed by greenlet during GC traversal.

In free-threading, some of the refs to objects held by the interpreter uses a special reference type called "deferred reference". Such references are used on some objects like functions and classes to reduce reference contention on the objects and they must be traversed by the free-threaded GC so that it doesn't free these objects prematurely.
Greenlet currently crashes on free-threading because greenlets hold suspended frames which on free-threading can contain objects with deferred reference count. The suspended frames needs to be GC traversed so that the GC doesn't free objects early causing a crash at runtime exit.

In 3.15+, I exposed _PyGC_VisitStackRef and _PyGC_VisitFrameStack in python/cpython#150767 which is used in this PR so that suspended frames of greenlets can be traversed by the GC. This fixes the crash and as a side effect also fixes the "A Cycle Of Greenlets Is A Leak" because suspended frames are now traversed so they can now be collected by the GC instead of being leaked. I have added tests on 3.15+ to ensure that suspended frames are gc collected and finalizers of locals are triggered properly.

[kumaraditya303]

ping @jamadden for review. I will create a followup PR to fix some of the known memory leaks of greenlet but that depends on this PR.

[jamadden]

Thanks for this! Today is the first time I've had a chance to look at it, and it LGTM. I can confirm it fixes the crashes on 3.15b2, both official from python.org and built locally (from MacPorts). I had a moment of panic due to a runtime linking error, but it turns out the needed symbol was only added in 3.15b2 and I was using 3.15b1.

I'm happy to merge this and add some appropriate notes to CHANGES.rst.