gh-135106: [crash repro] Randomly deposit objects in _Py_Dealloc

[colesbury]

Returns 1 a little less than 1% of the time in
_Py_RecursionLimit_GetMargin() to force the use of the trashcan mechanism.

This isn't intended to be merged. It's just to:

1. Demonstrate the crash
2. Test on a variety of platforms and configs from the CI

• Issue: test_weakref.test_threaded_weak_key_dict_deepcopy crash: merged objects should have ob_tid == 0 #135106

[markshannon]

I don't think this shows anything more than that randomly changing the stack limit causes crashes.

The trashcan mechanism relies on the margin being larger for frames higher up the call stack. This breaks that invariant.

You could try changing margin > 2 and margin >= 4 to much larger numbers to see if that provokes a failure.

[colesbury]

I don't think this shows anything more than that randomly changing the stack limit causes crashes.

The only thing that matters here is the condition to deposit objects -- there's no invariant to be broken. I've updated the test to make that more clear. It should always be safe to deposit an object conservatively (i.e., to treat the available stack as smaller).

Here's the equivalent test with the old trashcan behavior, which works without crashing:
https://github.com/colesbury/cpython/tree/gh-135106-test-old-behavior