Minor changes in how COMPLETING jobs and their nodes are handled.

If a node is down and not responding, don't bother to send a KILL_JOB
RPC to it. If that is the only node associated with a job, don't have
that job go through a COMPLETING state. It goes directly to a COMPLETED
state. Also preserve the NO_RESPOND flag associated with a node if its
state is changed via user request (e.g. scontrol).