feat: add Surf Inference to service directory

OpenAI-compatible LLM inference with per-token streaming sessions.
First-party MPP integration accepting Tempo USDC payments natively.

Models: Claude (Sonnet/Opus), Grok, Kimi, MiniMax, Qwen, GLM
Endpoints: POST /v1/chat/completions (dynamic pricing), GET /v1/models

Author: tenequm