Making jobs interruptible makes little difference to reducing the
workload / queue size of the runner, since it only applies in a pretty
specific case: when a new push happens to the same repository and branch
as the one job that is *currently* running. But *when* it happens, it
often gives us the worst of both worlds: CPU time is spent anyway
(maybe even 99% of the job) but we don't get a result from it.

I just saw a 34-minute job get interrupted after 25 minutes due to
another push, with nothing else queued. This means those 25 minutes were
completely wasted work. In the past I saw even longer wasted jobs. If we
already started running it, we might as well let it finish...

This reverts commit 36144a29.