Fix bug in job step allocation failing due to memory limit

This fixes a bug where a system is enforcing memory limits and
the job already has a step running on some of the nodes then
tries to start another step using some of those nodes. For example
wwith DefMemPerNode configured and the select plugin enforcing
memory limits, try:
salloc -N2 bash
$ srun -N1 sleep 10&
$ srun -N2 hostname
Without this patch, the second srun would fail instead of pend.