Add slurm GPU variables to GPU docs

Jobs using multiple GPUs should use SLURM_GPUS_ON_NODE to make effective use of all available GPUs.