Reminds of some experiments someone did a while ago as part of the
GIL removal attempts where the ref count integers are all kept in a
separate array. The intent there was to be able to do locking on
a single array rather than on individual decref cells.

This would solve the issue with having to jump around in memory
to decref all objects, but I'm not sure whether the overall win
would be a lot, since deallocation of the memory blocks typically
requires accessing the block itself as well (to update the block
chain list pointers), unless the memory allocator uses some
smart cache local block management as well (I believe that pymalloc
does, but could be wrong).

In any case, this sounds like a fun experiment for a GSoC student.
Perhaps the PSF could donate an AWS EC2 instance with enough RAM to
do the experiments.