sets new environment variable SLURM_CPUS_ON_NODE for use by LAM/MPI.
Also fixed bug in srun task distribution logic for block distribution
in heterogeneous cluster.

Otherwise signal all steps associated with the job (unless individual job
steps are identified).

signals all job steps, but not the job script itself).

   - fixes to help ensure slurmd uses the same key for shared memory
     on a restart (to avoid losing track of jobs)
   - slurmd only runs one launch thread at a time
   - fix bug in slurmd where multiple threads used same address space
     for connecting client address.
   - srun always sends SIGKILL to job step before issuing complete request
   - Changed short string for draining nodes to drng from drain.
   - srun default launch message timeout increased to 5s.

 o Modify srun(1) man page to reflect that, really, slurmd debug level
   is only allowed to be set up to 4.

to take effect.

scontrol command to use it.

and/or job step(s) will have their resources de-allocated and be killed.
A resource allocation will not be release unless no job steps are active
for at least InactiveLimit seconds. DPCS jobs will be subject to this
forced de-allocation if they remain inactive for an extended period of
time, which can get SLURM and DPCS back in sync if DPCS does a cold-start.

report an error message.

avoid highly fragmented resource allocations.
Add list of excluded nodes to job info dumpped and reported.
Fix how mis-matched RPC version number are handled. Let error code get
back to the API function.
Dump job state information upon each job's termination via plugin.
Re-issue incomplete write requests in job/partition state save.
Make slurmctld continue proper operation without any default partition
(gnats:317).
Add command/RPC to delete a partition.
Retry socket connection for slurmd/io.c as needed (gnats:253).

equivalents. For consistency make them return integer values (like before)
that are not valid characters (e.g. 0x100). This avoid the use of single
character values that we might want to preserve.

and --nodes options.

   on nodes relative to the current allocation.
 o srun no longer sends SIGKILL to job if one task is killed except
   if --no-allocate is used. (the job will otherwise be killed by
   the controller anyway)