Self-hosted LLM gateway

OpenAI-compatible reverse proxy in front of on-premise inference backends — rate limiting, GPU capacity management, and usage accounting for Elettra.