It seems that when multijob resources set up their subjobs they subscribe to
subjob notifications in such a way that they will never exit memory. In
particular, the multijob resource itself implements implements the notify
callback interface and is registered with the notification consumer manager for
each subjob, which ends up keeping it in memory (this step needs to be
proxied). The other issue is that upon destroy we only remove one of the subjob
notification consumers, which essentially leaves the multijob resource in
memory indefinitely.