error instead of doing an assertion failure.

  found in testing that new code:
- avoid draining node if stdout/err file can't be created
- correction in logic for >1 thread per core in setting a job's SLURM_CPUS_ON_NODE env var
- Fix for uneven task distributions on heterogeneous clusters
  (e.g. 4 CPUs on the first node and 2 CPUs on the second,
  a task_count=2 resulted in cpus_per_task=3, and previously
  allocated both tasks to the first node, now does one task
  per node)

 Mostly a matter of moving some logic to a new module.