bpo-44881: Integrate GC untrack into trashcan begin.

[nascheme]

For subtle reasons, PyObject_GC_UnTrack() function must be called before
Py_TRASHCAN_BEGIN(). There have been a number of bugs over the years
related to not doing this particular dance just right. Integrating
the PyObject_GC_UnTrack() call makes it harder to do things incorrectly.
That avoids some hard to find bugs (e.g. only triggered when object
nesting gets deep enough).

Extensions that still call PyObject_GC_UnTrack() explictly will
work correctly but the call is unneeded after this change. It would
still be needed for the extension to work correctly with older versions
of Python.

https://bugs.python.org/issue44881

[pablogsal]

Let's land #27678 first to include that change in this PR

[nascheme]

Just to be clear, I don't consider this PR a bugfix and it would not be appropriate to include in 3.10.

[pablogsal]

Just to be clear, I don't consider this PR a bugfix and it would not be appropriate to include in 3.10.

Yep, we are on the same page :)

[ambv]

We won't be merging this as of yet because an alternative solution is evaluated in GH-27738. I added the assert Victor wanted so that the benchmarks are fairer between the two approaches.

[github-actions]

This PR is stale because it has been open for 30 days with no activity.

[nascheme]

This PR is out of date and will not be required if GH-132280 gets merged. That PR integrates the trashcan mechanism into Py_Dealloc and also stops using the GC next/prev pointers to store the queue of trash. So, it removes the need to call PyObject_GC_UnTrack() at the exact right time.