Refresh DCGM exporter

[dholt]

Summary

• Refresh DCGM Exporter from stale Ubuntu 20.04-era pins to NVIDIA/dcgm-exporter 4.5.3-4.8.2.
• Refresh bundled default counters from the upstream tag.
• Adjust the Kubernetes DaemonSet manifest for the distroless image entrypoint and tighten container security context fields.

Validation

• Public GitHub Actions on final head a6242f6a4696d1c06304db988a123d22b457267c: PASS
• ansible-lint roles/ via the DeepOps role lint target: PASS
• ansible-playbook --syntax-check playbooks/slurm-cluster/nvidia-dcgm-exporter.yml: PASS
• YAML parse of modified defaults and DaemonSet manifests: PASS
• git diff --check: PASS
• docker manifest inspect nvcr.io/nvidia/k8s/dcgm-exporter:4.5.3-4.8.2-distroless: PASS for an OCI index with linux/amd64 and linux/arm64 manifests
• Refreshed CSV files compare exactly to upstream etc/default-counters.csv at tag 4.5.3-4.8.2
• Static OS compatibility audit: DCGM monitoring high findings reduced from 3 to 0
• Live single-node Ubuntu 22.04 GPU-server deployment: PASS
  • Ran ansible-playbook -i <inventory> playbooks/slurm-cluster/nvidia-dcgm-exporter.yml -e hostlist=all --flush-cache from a pinned DeepOps Ansible 10.7.0 controller virtualenv.
  • Play recap: ok=94 changed=31 unreachable=0 failed=0 skipped=30.
  • Post-run checks: docker.dcgm-exporter.service active/enabled, container image nvcr.io/nvidia/k8s/dcgm-exporter:4.5.3-4.8.2-distroless running, nvidia-smi healthy on two GPUs, and curl http://127.0.0.1:9400/metrics returned GPU metrics including clocks, temperature, power, utilization, and framebuffer gauges.
• PR body sanitization: PASS

Notes

• Upstream release: https://github.com/NVIDIA/dcgm-exporter/releases/tag/4.5.3-4.8.2