feat: provision AMD Developer Cloud MI300X model endpoint

[Ker102]

Goal

Provision the primary AMD Developer Cloud / DigitalOcean MI300X inference endpoint for nullstate red/blue agents.

Acceptance Criteria

• AMD Developer Cloud / DigitalOcean access confirmed
• ROCm or provider GPU image selected
• vLLM or SGLang serves an OpenAI-compatible endpoint
• NULLSTATE_LLM_BASE_URL works with python -m nullstate run examples/azure-public-blob
• amd-smi or rocm-smi evidence captured
• vLLM /metrics snapshot captured

Technical Notes

Keep Fireworks as a contingency, but the primary case-study route is self-controlled AMD GPU serving.