feat: collect vLLM and AMD GPU metrics artifacts

[Ker102]

Goal

Collect model-serving and GPU evidence for the hackathon case study.

Acceptance Criteria

• scrape vLLM /metrics before and after a run
• parse key counters into metrics.json
• capture raw Prometheus snapshots as run artifacts
• capture amd-smi or rocm-smi output when available
• document endpoint type: offline, managed, or AMD GPU-hosted

Technical Notes

Do not require GPU tooling for offline mode.