Improve handling of --gpus-per-task to make sure appropriate number of GPUs

is assigned to job.

Factor --gpus-per-task(gres_per_task) in --gpus (gres_per_job) count

Logic of _eval_nodes has appropriate handling for per job gpus, which
should be used also in case of GRES specified by --gpus-per-task. This
can be achieved by properly initializing gres_per_job if by
gres_per_task * num_tasks.

Bug 8133.