Send pack job's task IDs, node count, tasks_per_node and node_list
  from srun to slurmd as needed by MPI. This information is constructed
  by srun combining information from each individual job component's
  step information.
Set global SLURM_NODELIST, SLURM_JOB_NODELIST, SLURM_STEP_NUM_NODES,
  SLURM_TASKS_PER_NODE, SLURM_NNODES, SLURM_NTAKS and SLURM_STEP_NUM_TASKS
   for pack step.
Modify pmi2 plugin to launch a single agent for pack step rather than
   one per component.
Modify pmi2 and pmix plugins to gather global step information for
   pack steps.
Make task distribtion be configurable by step component (i.e. "block"
   on one and "cyclic" on another).

Required for pack step support

Enforce this assumption at configure time; if this is not true
there are some serious security risks that would result.

Previous commit removed this flag as part of Cygwin/Solaris cleanup.

Which then removes them from Makefile.in files after autogen.sh.

Cygwin removal previously done in ada78c4e, but the configure
script was defined.

Solaris is a dead platform; commits to remove last traces of it
will follow.

- gid instead of uid used in group_cache_lookup_job().
- needle->username is optional, and will lead to a failure
  in getgrouplist() when NULL is passed in.

Hat tip to Brian for discovering.

in salloc, sbatch, srun before process the command line options.
Bug 3499

plugin when constraining devices.

from the string.

correctly in sacct.

CID 177191.

Modify _open_as_other so that both _file_bcast_register_file
and _prolog_error can both use it. (That code was originally
extracted from _file_bcast_register_file, but tailored to
_prolog_error's requirements. Less duplication is good.)

In slurmctld:
  - Handle the lookup, and send results across.
  - Periodically flush the cache.

In slurmd:
  - Load value sent as part of prolog into cache
  - Have task launch use the new call instead to fetch
    the jobid-cached values when available
  - Remove old job-specific records from the cache as
    part of the prolog.
  - Ignore the slowly increasing memory footprint from the cache,
    as this is what the old implementation did anyways. It'll be
    O(users in cluster) at worst, and flushes on a reconfigure.
  - Simplify the slurmstepd communication path. group_cache_lookup
    will always succeed.

Left to do:
  - Convert other lookup locations in slurmd to the new approach,
    and remove the compatibility shim.

Bug 3322.

Loosely based on src/slurmd/req.c. But with major alterations
in preparation for use with PrologFlags=SendGIDs.

Convert slurmd to use a compatibility shim in here for now.

Bug 3322.

The slurmd is always sending the ngid / gids over
the pipe, so the setgroups() call is all that matters
here. Move it out to _drop_priviledges instead.

Bug 3322.

Patch was broken by a rebase before pushing. Fixing as a clean patch.

This reverts commit b0838485.

Regression caught my slight oversight from 88d5a801.

Bug 3322.

Always false in all supported plugins, except for BGQ.
Prior commit removed all calling paths.

mpi_hook_client_single_task_per_node() is always false,
except on BGQ systems. Move special handling code into
an #ifdef so it can be readily identified and removed
when BGQ support is retired.