Without this change, when the slurmctld daemon is reconfigured and
jobs are in completing state, then all nodes allocated to the job
are resent TERMINATE_JOB RPCs. After this change, only nodes which
have not processed the TERMINATE_JOB RPC are sent another one.

gpu_device global data structure.

Downing the node will kill all jobs allocated to the node, very bad
on something like a BlueGene system

multiple memory operations which are not necessary.

If slurm.conf defines fewer of a gres than gres.conf AND gres.conf
defines specific CPUs to be associated with the gres, this patch
prevents the bit_overlap call being made with bitmaps of different
sizes, which would cause an abort.

Also modify the epilog script to use this new output field

Timing changes due to faster job step launch

This can be especially useful to schedule GPUs. For example a node can be
associated with two Slurm partitions (e.g. "cpu" and "gpu") and the
partition/queue "cpu" could be limited to only a subset of the node's CPUs,
insuring that one or more CPUs would be available to jobs in the "gpu"
partition/queue.

This happens when a job has multiple partitions and
priority/multifactor is NOT in use

fix of the following problem:
if a node is excised from a job and a reconfiguration(e.g., update
partition) is done when the job is still running, the node will be left
in state idle but not available any more until the next
reconfiguration/restart of slurmctld after the job finished.

to be edited more.

to be running in the calling program.