fix cons_tres task layout with gpus-per-task

This fixes a potential task layout problem. Specifically if
a cluster allocates by core and each core contains 2 CPUs and
a job requests 3 tasks and one core on each of two nodes is
available and the first node has one GPU available and the
job requests --gpus-per-task=1 then without this patch
cons_tres would try to put 2 tasks on the first node (one
per CPU). This adds a check of the GPU count in order to
prevent that. Observed sportatically when running regression
tests39.[10-15] in immediate succession on a Cray system
or a system configured with Epilog that takes a few seconds
to complete.