Make sure count used is that for a single GPU rather than a total MPS
over multiple GPUs

Coverity CID 190461

This would be triggered by the gres.conf on the compute node having a
higher count of GRES defined with "File" parameters than the node has
defined in slurm.conf.

This also adds documentation and some tests. The actual scheduling
logic is still not written.

No new fields added, just duplicated the save/restore for v19.05

# Conflicts:
#	doc/man/man5/slurm.conf.5
#	src/plugins/sched/backfill/backfill.c
#	src/slurmctld/job_scheduler.c

Coverity CID 190390, 190393, 190397

Enabling this alters the scheduling sorting algorithm so that hetjob
components belonging to the same hetjob are attempted to be scheduled
consecutively (not interleaved with other jobs or hetjob components).

Bug 5579.

AccrueTime limits.

Bug 6002


Co-authored-by: Felip Moll <felip.moll@schedmd.com>

Coverity CID 190390, 190393, 190397

Fix for Coverity CID 190461

Add small note to documentation to clarify how to set different SlurmUser for
slurmdbd and ctld.

Bug 6306

This fleshes out logic so scontrol can be used to update a node's GRES.

Previous logic was failing to report errors in node's GRES update
requests.

No change in logic, improve some comments and logging

This is only used to define the slurmctld_comm_addr symbol in allocate.h.
Variable definitions do not belong in headers. Rather than relocate the
struct declaration, replace the struct which only holds a single port number
with a static variable serving the same purpose in allocate.c, and the
extern in this same location so srun.c can find it as well.

Replaces the Linux SO_PEERCRED calls.

This improves handling of job requests for GRES with zero counts.
Previously if gres/foo had a count of zero, then "srun --gres=foo ..."
would generate an invalid TRES error, while "srun --gres=foo:1 ..."
would just pend. Now they both pend.

Regression introduced in ccafaf7b.

Replace the (was always set to true) check_current_dir argument to
search_path() with check_cwd_last boolean to indicate if cwd should be
evaluated first (needed by sbatch) or last (salloc/srun).

Bug 6271.

Without this patch the Gres count might be reported as "NAME:0G" rather
than just "NAME:0".

No functional change intended.

Kill running/pending job is allocated GRES and that GRES has a "File"
    configuration, and the GRES count changes.
Also generally better handle changes in GRES count (avoid some bugs
    in GRES data structures).

(Name change presumably missed due to the typo.)

This properly sets up the node's GRES data structure for gres/mps
with CLOUD based nodes. This sets the counters appropriately for
nodes which are not up when the slurmctld restarts.

This fixes test39.18 to work with slurm logic corrected in
commit 992e393f

Coverity CID 190597