The launch for pack-job 1+ fails in api/step_launch.c slurm_step_launch_add()
  right now, but the logic in srun seems OK.

The job allocation response does not include the time limit and
  it would not be used later in the srun logic anyway.

Rather than the launch plugins directly access srun's "opt" options
  data structure, pass a pointer to the options specifications for
  that particular portion of a heterogeneous job.

Fix some parsing issues with respect to memory management, avoid
  memory over-writing and double free
Create one step per pack job (no work on launch yet)

Needed to copy some strings to avoid memory corruption

Fix problem reported by Clang

Validate allocation against request for pack jobs
Set some pack job environment variables

Without this, terminating srun with ctrl-c will leave job
  allocation around

srun --test-only option functional for pack jobs

Document which srun-specific options will be propagated between
  job steps rather than reset. Also propagate checkpoint options
  for consistency with sbatch.

It is needed by the perl APIs

This is just phase 1 of the parsing changes required for srun
  to support pack jobs

This will route a heterogeneous job submitted to multiple clusters
  to the one cluster able to start all compoments fastest.

salloc was failing to report the "Granted job allocation #" message
  and on termination was only deleting the last pack job component.

Fix problems reported by valgrind, several memory leaks and one
  memory reference after free.

This is needed to support requeue and other operations on the
  pack job as a whole

Make sure the pack_job_id values match, just in case something
  bad happended with the job records. If they don't match, log
  an error and skip the records.

Always start the script in the correct component's allocation
  and get the script from the corrrect location (could work before
  but dependent upon race condition).

If num_tasks == NO_VAL but min_cpus != NO_VAL, then num_cpus gets
reset to NO_VAL, which results in select/cons_res getting stuck in
a very slow loop (counting up to NO_VAL).

There are a couple more left to handle

Copy the strings into job request descriptor rather than just copying the
variable pointers to avoid over-writing contents

Test for unset job node_bitmap as the job may not yet be scheduled