If we create a new cgroup but do not manage to move a PID into it, then
we ended up leaking the inotify handler. This is because the
IN_DELETE_SELF will never actually fire at a time that we wanted. We'll
get an event either way when there is a process because the removal of
the last process is a change in itself.

As such, only listen for IN_MODIFY and unref the inotify watch ourselves
in case we succeed in deleting the cgroup again (or, in case someone
else managed to race with us and it is gone already).

Fixes: #6