Containerized vllm and rhaiis

[csibbitt]

• Adds nvidia Container Toolkit
• Configures Container Device Interface
• Run vllm from a container
  • Frees us from the python version management
• Updated flavor specs from testing
• Also works with RHAIIS
• BONUS: Idempotence checks for glance image and floating IP that got accidentally dropped on a previous refactor

[csibbitt]

This PR moves the workload execution into a container. By default it runs upstream vllm / TinyLlama, but with the right options, it will run RHAIIS / Llama-2.3-1B instead.

I tested both workloads on my nested-passthrough setup, upstream vllm by applying the AnsibleTestCR.yaml as-is, and RHAIIS by adding the variables shown below. They are both working fine.

Variables for RHAIIS:

    gpu_validation_workload_container_image: "registry.redhat.io/rhaiis/vllm-cuda-rhel9:3.0.0"
    gpu_validation_model_download_registry_username: "|3c5aa7e0..."
    gpu_validation_model_download_registry_password: "eyJhbGci..."
    gpu_validation_workload_cache_mount_path: "/opt/app-root/src/.cache"
    gpu_validation_model_name: RedHatAI/Llama-3.2-1B-Instruct-FP8
    gpu_validation_workload_userns: "--userns=keep-id:uid=1001"

[MiguelCarpio]

/lgtm