Ouch! We totally forgot about sub interpreters during the
reimplementation of the atexit module. 

In my opinion it's a big bug for systems like mod_python and mod_wsgi.
However it's too late to fix it for 3.0.0. Perhaps we can fix it in
3.0.1. We always knew that Python 3.0 is going to have some bugs that
can only be revealed by 3rd party code. I've added Barry to the nosy
list in order to get his opinion as release manager. 

Do you have a suggestion how the problem should be solved? I'd store the
atexit_callback information in PyInterpreterState instead of the module.
Maybe Martin's PEP 3121 could come into play to rescue us.

By the way PEP 3121 ... does the new module initialization code play
nice with multiple sub interpreters?