Random columns in CSV

[benoit-cty]

Describe the bug
Hi,
I've recently encountered an issue with successive measurement recordings in CSV format using CodeCarbon. When running the same measurement multiple times in a row, the CSV format appears to change unpredictably: the gpu_count and gpu_model columns intermittently appear and disappear without any clear reason.
This inconsistency triggers the warning: "The CSV format has changed, backing up old emission file." As a result, multiple files are created: each previous file is backed up with a .bak extension, and a new file is generated.
Is this a known issue? If so, is there a way to resolve or work around it?
Environment Details:
• CodeCarbon version: 3.2.3
• OS: Windows-10-10.0.26200-SP0
• CPU: 11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz
• GPU: None
Thanks in advance for your help!

To Reproduce
Steps to reproduce the behavior:
successive measurement recordings in CSV format

Expected behavior
Consistent columns

[cianc]

This is probably happening because we drop empty columns here:

codecarbon/codecarbon/output_methods/file.py

Line 107 in 01c2fc5

df = df.dropna(axis=1, how="all")

We could possibly remove this logic since that futurewarning is gone in pandas 3.0.0, but I think we should also try and figure out why those gpu fields keep flipping between None and non-None values.

The environment details above say "GPU: None" - is that definitive, do we know what values are showing up in the csv output when it's not None?