> (*) I made the GC state per-interpreter: commit 7247407c35330f3f6292f1d40606b7ba6afd5700 (Nov 20, 2019)

FYI, this was done by me in an earlier comment which we ended up
reverting.  Later you basically un.reverted that.

> The bug is that a C function object (_sre.compile) is created in an interpreter, tracked by the GC list of this interpreter, and then it is destroye and untracked in another interpreter.

FWIW, at one point I had a branch that supported sharing read-only
Py_Buffer data.  When the receiving interpreter was done with it I'd
call Py_AddPendingCall() to schedule the cleanup in the "owner"
interpreter.  However, this only worked because I kept track of the
owner.  Adding that pointer to every object wouldn't be feasible but I
suppose there are other things we could do that wouldn't be super
inefficient, like don't worry about it for the main interpreter, use a
hash table (Victor's idea), borrow some of the bits of the PyObject
head to store a flag or even an index into an array (if there are only
a few interpreters), or even make the allocator per-interpreter and
then extrapolate the interpreter based on the object's address.

Regardless, it is still much simpler to make all objects per-interpreter.