Use --ntasks to specify cores instead of --cpus-per-task:
SLURM_CPUS_PER_TASK now clashes with SLURM_TRES_PER_TASK

The old behaviour of loading bork group specific settings can be enabled
by passing --bork.
The short setting -B has changed its meaning from --nobork to --bork.

Currently supports only sharing with users in the same UNIX group

In some cases NumTasks would be set to 0 causing resource allocation errors

Instead of trying to escape the string a second time use srun as
interpreter of the script.
Not sure if this will cause srun to run properly or if it will instead
run as bash. This will need testing with multiple nodes which was why we
used srun in the first place.

closes #1

--nobork is also available to disable this behavior if necessary

This introduces a backwards incompatible change that causes memory
availability checking to be unreliable.
Basically without h_rss or h_vmem limits we don't have a reliable way to
know how much memory is going to be used/reserved.

Before, chaining multiple commands in the same line:

  cmd1 ; cmd2

would result in unintuitive errors due to running:

  srun cmd1 ; cmd2

instead of both commands inside the same 'srun' invocation.

Most SGE setups set memory through h_vmem. This however sets virtual
memory, not 'real' memory. Some software tries to reserve more memory
than necessary which leads to OOM-kill events.
Using h_rss works around this problem. However h_rss is not used in the
wild as often as h_vmem.

In this mode the job script is still processed and expanded according to
the selected queuing system but most options won't have any effect since
execution will be done locally and without submission.