gpu-utilization.sh does not properly collate multiple GPUs per pod

[starpit]

This can result in apparently k samples per timestamp, if the pod has access to k gpus. We will need to discriminate the signals by GPU. Maybe using the bus id attribute from nvidia-smi.