Support Red Hat NVIDIA Container Toolkit path

[dholt]

Summary

• add Red Hat-family support to the native nvidia_container_toolkit role
• route RHEL/Rocky 8+ hosts through NVIDIA Container Toolkit instead of legacy nvidia.nvidia_docker
• refresh RPM airgap guidance for EL9/Rocky-era CUDA and Container Toolkit repositories

Validation

• git diff --check
• YAML parse of changed Ansible files
• ansible-playbook -i <artifact inventory> playbooks/container/nvidia-docker.yml --syntax-check -e hostlist=all -e ansible_python_interpreter=/usr/bin/python3
• ./scripts/deepops/ansible-lint-roles.sh
• Live Rocky Linux 9.7 validation on Colossus ipp1-1946 (A100X x2):
  • Docker install passed after --flush-cache; first run exposed stale Ansible fact cache from a reused inventory alias, fixed in private maintenance harness commit a66fb18
  • playbooks/nvidia-software/nvidia-driver.yml installed NVIDIA driver 580.159.04; the role updated the Rocky kernel to 5.14.0-687.10.1.el9_8.0.1 after exact headers for the provisioned kernel were unavailable, rebooted, and nvidia-smi passed
  • playbooks/container/nvidia-docker.yml selected nvidia_container_toolkit on Rocky 9.7 and skipped legacy nvidia.nvidia_docker
  • idempotence rerun: ok=16 changed=0 unreachable=0 failed=0 skipped=4
  • nvidia-container-cli --version: 1.19.1
  • /etc/docker/daemon.json sets default-runtime: nvidia
  • docker run --rm --gpus all nvcr.io/nvidia/cuda:13.0.2-base-ubuntu24.04 nvidia-smi saw both NVIDIA A100X GPUs with driver 580.159.04

Artifact directory: /home/dholt/deepops-release-artifacts/26.07/20260528-161023-container-toolkit-rhel-live-validation