NVIDIA 6 added differentiation between FB and BAR1 memory - use nvidia_gpu_mem for FB Memory *only* since its definition fits better with the usage prior to BAR1 memory.
Memory used percentages were being concatenated into a single string if multiple GPUs were present, e.g. both using 30% would print mem0.value 3030 mem1.value
TODO 1: Follow multigraph suggestion from Flameeyes to look into multigraph
plugins (http://munin-monitoring.org/wiki/MultigraphSampleOutput),
in order to reduce the amount of round trips to get the data.
TODO 2: Put warning and critical as vars in config with sensible defaults.
TODO 3: Add additional output options for nvidia-smi only available for
professional GPUs.
FIXME: Possible bug in lines 87-91 of amd_gpu_