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.

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

Slight hit from heap based strings instead of stack, but worth
it to remove sprintf() call in lightly-used debug path.

Rather than reworking to remove the sprintf(), just remove this now.

This does change the output slightly in one_liner mode, in that
there will no longer be double-spaces between elements that would
have been broken up on successive lines.

The first case here is even more fun... let's just let
xstrfmtcat() sort this all out instead of manually determing
the required buffer size.

Functionality is still available if required, but has been deprecated
since 17.11, and has been removed on master before 19.05.

Un-document is now lest people start trying to use it.

Bug 5712.