are signalled to shutdown cleanly. Wait for each to finish to avoid too much
busy work at once; this may need revising if it turns out to be too slow with
a larger number of VMs (e.g. signal/delay/signal/delay/... then wait for
shutdowns), but let's avoid making it more complex unless we know it's needed.
Based on a diff from abieber@, discussed with mlarkin@ aja@ rpe@, ok rpe