Tracking not supported for NVIDIA Blackwell GPUs

[raphischer]

Issue:

When testing CodeCarbon on an NVIDIA DGX Spark, I noticed that the tracker cannot be correctly initialized. When debugging into that, I noticed that calling pynvml.nvmlDeviceGetMemoryInfo in codecarbon/core/gpu.py throws an pynvml.NVMLError_NotSupported, probably due to the shared memore architecture of the NVIDIA Blackwell GPU (internally read as NVIDIA GB10). nvidia-smi currently also lists N/A for some static information.

Solution:

Catch the error and instead return negative default values. As far as I see, this information is only statically stored and thus does not negatively impact the downstream functionality of CodeCarbon.