Move defines above function.
Remove TODO comment.

Bug 6635

Clarify NVML-related sub-tests.
Add message indicating convenient test feature.

Clarify function description comments.
Change a fatal to xassert, since it's really a development check.
Add a fatal if NVML happens to give us an empty CPU affinity array.
Update test 38.19 sub-test 0.14 to expect this fatal (in test prog).

Bug 6635

As tmp and the underlying wait_all_nodes fields are unsigned this
can never be negative.

CID 197666.

CID 197667.

CID 197668.

CID 197670.

Do not set job's per-node minimum memory on error
Bug 5562 Comment 27

This reworks all CLI option parsing for salloc/sbatch/srun into a central
set of functions. This is designed to allow a future set of commits to
inspect, and potentially modify or unset, the set of options as specified
by the user through a consistent string based set of functions.

Bug 3745.

Invalid --cpu-freq options now cause srun to exit(-1), rather than
continuing with the job submission. The test had been designed to
assume the job would still be submitted, so modify it to match
the new stricter parsing.

Have slurm_process_option() call exit(-1) on errors. Most parsing
functions will already do this, but that should be revisited at a
later point in time.

Remove the options argument now that no calling paths need it.

Note that opt_string here is alphabetical *by the option name*,
and not by the val any longer.

Move handling for parallel debugging into _opt_verify(), and
explicitly call through slurm_process_option() to override the
default settings.

It's just an unused copy of nodefile at this point.

Drop _get_int() as this was the last calling location.

All env vars are now handled through slurm_process_options().

There is a mention in a3449c0b that it was pulled out of
slurm_opt_t previously because the Perl API needs it; I suspect the
path requiring it was removed with either BG/Q or Cray ALPS, and
this is no longer a strict requirement. At least there are no
references to mpi_type left there.

Drop the cpu_bind_type_set bool from srun_opt_t; this can already
be deduced from whether cpu_bind_type has any bits (aside from
CPU_BIND_VERBOSE) set.

Move warning about excessive values for --threads into arg_set_max_threads,
but soften the language since the limit has not been enforced for a while.

Do not reset this each pass. Especially since the only effect this option
has is to bump pmi_server_max_threads() around, and that doesn't not know
about HetJobs.

Move MAX_THREADS to slurm_opt.h, and rename to SRUN_MAX_THREADS to avoid
collision with the same macro name elsewhere. This should probably be made
into a configuration value at some point.

Drop relative_set in favor of direct test against NO_VAL.
Zero is also a valid value to have set this to - change two tests
to only check against NO_VAL.

Note that this argument's behavior is different from --wait in sbatch.