tests were change appropriately. This lets one fake a Linux cluster
without all of the BGL-specific logic.

and FastSchedule=0 and the cluster is heterogeneous.

Not critical, but a loose end in found in tracking memory leaks.

 uninitialized variables (thanks to valgrind).

  (was continuing search if job specified node count, but not geometry).

No changes in logic.

listed twice.

used on Blue Gene systems. This avoid a redundant message for each and
every back-end node.

hostlist_ranged_string instead of one line per node.

get set DOWN rather than just the first one.

It will be commonplace on Blue Gene system.

Still looking for reason the plugin adds 15MB to slurmctld's size
(compared with select/linear).

 for use in job and partition print modules.

to attempt scheduling every pending job in priority order. Don't
stop scheduling for a SLURM partition after first job scheduling
failure as many different job parameters are considered: geometry,
rotate, conn-type, node-use.

Return number of bytes written on sucess rather than SLURM_SUCCESS.