feat: axum OpenAI-compatible REST server (--features serve)

Endpoints: /health, /v1/models, /v1/chat/completions, /v1/completions
ModelRouter dispatches to GPT-2/OpenChat/future engines.
serde_json for request/response serialization.

cargo run --bin serve --features serve --release
curl http://localhost:3000/v1/models

Tested: health ok, models list returns, chat/completions wired.
No model weights loaded yet — needs register_gpt2(weights) at startup.

https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK

Author: claude