Minimize preempted jobs for configurations with multiple jobs per node.
  Previous logic would preeempt every job on node allocated to pending
  job.
bug 2906

Fix for core selection with job --gres-flags=enforce-binding option.
    Previous logic would in some cases allocate a job zero cores, resulting in
    slurmctld abort.
bug 2808

Add some indentation so that GRES topology-specific information
  logged is more readable.

A job allocation selecting nodes and no cores/CPUs could write
  off the end of arrays and corrupt memory. Now to figure out how
  the logic reached this point in the first place.
bug 2808

In commit b8190e5d many places that were mean to be pending step ids
were changed to be extern_step id.  The main problem was when we came up
with the idea of the extern step we reused -1 (INFINITE) for the id.  So
pending steps also appeared to be extern steps as well.  Hopefully this
fixes the situation.

Bug 2907

Start power save thread only after the partition information is read
  in order to avoid trying to interpret the SuspendExcParts configuration
  information before the partition information is available, which would
  result in a slurmctld abort.

Do not try to access part_list variable (partition list pointer)
  if not yet initialized. Return NULL pointer rather than aborting
  with NULL pointer.

bug 2900

Don't register newly found buffers that are less than OtherTimout
  old to avoid possible duplicates.

This hardens the code with respect to a race condtion if the
  slurmctld restarts and a burts buffer creation for a job is
  in progress. Eliminate the possibility of a duplicate job
  allocation record.

No change in functionality, just moved function call and added
  comment

delete_step_records which would delete the steps without the killing flag
set.

What this does is treats the extern step like a normal step on exit.  It
doesn't appear the original code is needed anymore and this simplifies
the code.

The select_cray change is relevant since the add is needed only when
killing the step as that is the only place _internal_step_complete isn't
used.

functions.

Before it was showing it as TBD since pending steps and the extern step
have the same stepid.

What this does is set the state earlier to match a normal set.

Remove the unneeded _send_pending_exit_msgs.  There is only one task and
we have the message for it, so don't worry about that one.

Most important, wait for the other slurmstepd's to send their message,
otherwise they could be lost on the other end.

Only execute the DataWarp real_size function if there is a job burst buffer. Calling the function if the job only references persistent buffers generates an error that is not useful

Wrong argument type

Preserve variable resp_msg for use in error message and use a different
  variable for temporary storage.

Fix gang scheduling and license release logic if single node job killed on
    bad node. Notifying gang and releasing licences is normally done when
    the epilog completion happens, but if the node(s) assigned to a job are
    all down, that does not happen. This results in the licenses being
    reserved indefinitely and the gang scheduler being left with a bad
    (old) job pointer that can result in various failure modes
bug 2867

Add hotels. Other minor changes.

anyway to attempt to log the backtraces of the potential
unkillable processes.