feat: add OpenAI-compatible LLM support (LiteLLM, retries, smoke test)

[richrz]

Summary

Adds an OpenAI-compatible LLM adapter to Codebuff.

Key changes

• Drop node-fetch; use native fetch (Node 18+).
• Add timeout + retries with backoff.
• Support Azure api-version + OpenRouter headers.
• Pass through tools/tool_choice.
• Harden SSE streaming parser.
• Add smoke:llm script.
• Add .env.example.

How to test

1. Copy .env.example → .env and fill values.
2. Build:
   pnpm build

   or

   npx sucrase ./backend/src/llm -d ./backend/tmpdist/llm --transforms typescript
3. Run:
   pnpm smoke:llm
4. Expected output: "Hi! 👋 How’s your day going?"

Notes

Works with LiteLLM, OpenRouter, Azure OpenAI, or any OpenAI-compatible API.