[collectd 6] gpu_sysman: change GPU metric labels to resource attributes

[eero-t]

ChangeLog: gpu_sysman: change GPU metric labels to resource attributes

Changes to Sysman plugin:

• Move current GPU attributes from metric labels to resource attributes
• Include additional GPU attributes recommended by the OTEL spec

Changes elsewhere:

• resource.c: WA for GCC's warning bug when building gpu_sysman_test.c with more extensive GCC warnings

Split from #4267.

[eero-t]

Hm. On the two HWs / kernels I just tested, one did not report any FW info, and the other reported 2 of them. This PR reports info for just the last FW, as OTEL spec seems to assume there only being single FW => I'll file bug against OTEL spec.

(FW info support is almost half of the changes in PR.)

[eero-t]

Prometheus adds exported_ prefix to job & instance labels, which can be remapped with serviceMonitor:

apiVersion: monitoring.coreos.com/v1
kind: ServiceMonitor
metadata:
  name: collectd
  namespace: monitoring
spec:
  endpoints:
  - metricRelabelings:
    - sourceLabels: [exported_instance]
      targetLabel: pci_bdf
    - action: labeldrop
      regex: '^exported_.*$'
  selector:
    matchLabels:
      app: collectd

This is in draft state because resource attributes make querying metrics from Prometheus much more complicated.

EDIT: In most cases I think one would be showing data from target_info separately, so maybe that is not so bad. Apparently there's also some activity on Prometheus side to help with this.

But if one does wants any of the resource attributes with the queried metrics, one needs to:

• specify each of the labels separately in the query
• for each metric

I.e. instead of getting all metrics for given $NODE from $SERVICE with query like this:
{node="$NODE", service="$SERVICE"}

One needs to now query each of them separately with a query like this:
collectd_gpu_sysman_temperature_celsius{node="$NODE"} * on (exported_instance) group_left($LABELS) target_info

Which gives items like this (when LABELS=model,pci_bdf,pci_dev):

      {
        "metric": {
          "container": "collectd",
          "endpoint": "http",
          "instance": "10.32.130.125:9103",
          "job": "collectd",
          "location": "gpu-max",
          "model": "Intel(R) Iris(R) Xe MAX Graphics",
          "namespace": "monitoring",
          "node": "testnode",
          "pci_bdf": "0000:0a:00.0",
          "pci_dev": "0x4905",
          "pod": "collectd-d9m6z",
          "service": "collectd"
        },
        "value": [
          1713363392.117,
          "30"
        ]
      }

[eero-t]

Looking at the OTEL HW spec: https://opentelemetry.io/docs/specs/semconv/system/hardware-metrics/#common-hardware-attributes

It seems that all HW devices should provide also id and name attrributes, in addition to similar GPU model and serial attributes?

FYI: I've raised these and the earlier noticed issues in following OTEL spec ticket: open-telemetry/semantic-conventions#940

[eero-t]

Updated code according to feedback given in OTEL bug tracker, which dealt with the issues I raised above.

Will keep this still as "Draft" though, until #4299 is merged, the discussion in OTEL ticket has winded down and I've tested this a bit more.

[eero-t]

Questions answered in OTEL ticket and code changed to match => dropped Draft status.

Rebased GCC warning WA & FW attribute commits to be last ones, in case they should be dropped from this PR.