The test actually wasn't checking interrupt counts at all. Fixing it
required some other changes:
Make sure all harts get to run
Add some retries, since on a loaded machine against spike both harts
might not get to run, even if you give spike a generous amount of time
to do so.