the old_job_allocation response so as to conform with job step's specification.
For example, remove nodes that are excluded (-x), set the node count to
that requested (-N, eliminate nodes as required), etc. This requires
modifying the node list including the slurm_addr, and cpu_count arrays.

   allocation of a job step.

 o Move checks for launch or exit timeouts from IO thread into msg.c
 o Added new "undocumented" options --max-launch-time and --max-exit-timeout
   for use in testing.
 o Do not wantonly terminate IO thread when job moves into state
   TERMINATED, only exit on FORCETERM. (this allows collection of error
   messages from slurmds)
 o Small amount of code cleanup in io.c, signals.c

 o Don't print exit status as an error if user generated signal that killed it.
 o Set srun exit status appropriately.
 o Accept more environment variables to initialize options.
 o Other minor fixes.

   slurm_cred_t job credential.

 o src/srun : move report_{job/task}_status into job.c
              report task exit on one line per msg (instead of per task)
              add new task state for abnormal (nonzero) exit.

 o src/srun/ : srun code cleanup.
   Do not set stdin nonblocking as this has a side-effect of setting stdout
   and stderr nonblocking too.
 o src/slurmd : code cleanup. Terminate if read of public key fails.
 o src/common/credential/signature_utils* : code cleanup

 o srun : fix problem with loss of output from stdout stream.
          When stdio buffering is on, fprintf/fputs do not return errors
          when the buffers are apparently overflown (?), setting unbuffered
          IO seems to fix the problem.
 o slurmd/io.c : some slurmd io cleanup

 o src/srun   : various signal handling fixes, retry task launch req on failure

   creation functions into allocate.[ch]
 o src/common/xsignal.[ch] : xsignal returns old signal handler
 o src/slurmd : remove list of running threads - use number of active
   threads instead. (terminating slurmd sleeps until all active threads
   have vanished)

 o Only terminate IO thread if IO is done or job is in FORCETERM state

pthread_kill(tid, SIGHUP) to knock loose the hung poll()

   affect number of tasks in batch job steps.

TotalView (it was intercepting the SIGTERM and reporting srun's tasks
in addition to the user's tasks - very confusing).

   (to match documentation)
 o Use access(2) to determine if user has local access to file
   (instead of brute force examination of stat buffer)
 o issue warning if local command cannot be found with --batch

   srun/signals.c
 o added experimental --join option which allows attached srun
   to send signals and stdin to job. (stdin not working yet)
 o other srun and slurmd code cleanup

   calls never block) This is only used in slurmd/mgr.c right now,
   so that log functions called in io.c do not deadlock IO thread.
 o cleanup in srun/srun.c

   Use constant size buffer for stdin.
   slurmd/fname.[ch] : Pre-open and truncate all unique output filenames
   slurmd/io.c       : make call to truncate all output files.

 o slurmd: if there is room in cmdline, change argv to show
   what job a slurmd is managing.

   support documented uses of --input,output,error)
 o Allow premature termination of srun (i.e. when IO is not yet
   complete) by insistent user.
 o added xstrfmtcat() and xmemcat() to xstring.c

   still remain.
 o cbuf.[ch], list.[ch] : Checked in new revisions from lsd-tools
 o srun/io.c,msg.c : added io_thr_create(), msg_thr_create() to simplify
   code in main().
 o Other minor fixes/changes

initiation.
Establish limit of MAX_TASKS_PER_NODE and add matching error code.

for reporting lack of activity, update task and node state in a couple
of places where missing.

 o do not kill job allocation while still waiting on IO

Enhanced error logging.

 o rewrite of slurmd.c -- extensive changes therein
 o launch msg contains port numbers instead of slurm_addrs

Fix comment typos.
Add LLNL disclaimers. --No changes to any logic

to use.
Srun to use slurm_addr's from allocate RPC except in the event of no allocation
(-Z option), where it generates slurm_addr's itself.