This fixes a problem if a job has a gres_per_job value set and
multiple required nodes (by name). Previous logic could block
use of later nodes if the gres_per_job value was reached by
earlier selected nodes.

Previous logic would fail to skip over comma when gres name is
followed immediately by a comma and no count
Also remove logic to set pointers that were not used later

bug 5743

with some configurations and given a race condition test39.7 would
fail. This corrects some bad logic (failing to update CPU count when
cores are removed from consideration)

bug 5745

In some cases test1.88 failed due to specifying a node count range,
"-N 1-6".

Used "<" instead if "<<"

changes only apply to non-standard task distribution options:
serial, lln, spread, and busy

bug 5669

Bug 5711
Bug 5741

This reverts commit 204da0b6.

This reverts commit 450cd6c8.

Bug 5711

This will be re-reverted in 19.05 with other code changes we didn't feel
comfortable dealing with in 18.08.

to --cpu-bind

Continuation of a57fafb2

No longer just used when using profiling.

ic/ofed, fs/disk, fs/lustre, pages, vmem

Bug 5702

The select_p_job_test() function is called once for the cons_tres
plugin with all available nodes rather than multiple times based
upon incrementally adding nodes based upon weight. This change
modifies the _eval_nodes_busy() and _eval_nodes_spread() functions
to use low node weights first and incrementally add higher weight
nodes as needed

Cores can be removed from availability in order to co-allocate CPUs and GPUs.
This change updates the available CPU count on a node when cores are removed
from availability by clearing bits in the core_bitmap.

Log not just the job ID and node name, but core_bitmap
and sockets to be avoided.

"Rebood" changed to "Reboot"

This throws 0c3133a2 out, as that'd already been handled here previously.
(And thus we do not add a separate NEWS entry for 18.08.1 on this issue.)

The backup slurmctld can end up without an agent thread if
read_slurm_conf() takes too long to complete. Forcing the agent_init()
here again ensures it is restarted if not currently running.

This was fixed upstream in c1a537db.

Bug 5633.

"Generalized Hypercube - A Topology Plugin" was accidentally given
a 50 minute slot instead of 30 minutes as it should have been. Carry
the fix forward through the schedule, and cede the extra time to Google.

Bug found by clang