CUDA 12.9 invalid metric names

[ec-jt]

Fix exporter crash with CUDA 12.9 by sanitising invalid metric names (e.g. fields with [us]) to ensure Prometheus compatibility.

[utkuozdemir]

I thought I already handled the metrics ending with [us] in this PR - can you share one of the metric names that fail here?

[Harry-zklcdc]

The reason seems no the suffix of '[us]'.

I noticed that the metric inforom.checksum_validation can cause longer printing times, leading to crashes.

The metrix was added at NVIDIA Linux Driver 575.

The following is an introduction to this metric by nvidia-smi --help-query-gpu

"inforom.checksum_validation"
Inforom Checksum Validation information.

[utkuozdemir]

The reason seems no the suffix of '[us]'.

I noticed that the metric inforom.checksum_validation can cause longer printing times, leading to crashes.

The metrix was added at NVIDIA Linux Driver 575.

The following is an introduction to this metric by nvidia-smi --help-query-gpu

"inforom.checksum_validation"

Inforom Checksum Validation information.

Should this be a separate issue then? Because it doesn't seem to be connected to the OP's issue.

[Harry-zklcdc]

Should this be a separate issue then? Because it doesn't seem to be connected to the OP's issue.这应该是一个单独的问题吗？因为它似乎与原帖的问题没有关联。

Because I find this problem when using NVIDIA Linux Driver 575.x.x (CUDA 12.9)/NVIDIA Linux Driver 580.x.x (CUDA 13.0) at 4090D/A10. It may cause NvidiaGpuExporter crash.

And It find that NvidiaGpuExporter would be OK after remove this metric.

Actually, this problem occurs when the GPU is under high load.