Critical fix for concurrent performance:
- Wrap sentence_transformers.encode() in asyncio.to_thread()
- Prevents CPU-bound sync operations from blocking FastAPI event loop
- Ensures server stays responsive during embedding processing

Code quality improvements per maintainer feedback:
- Remove duplicate async embedding functions
- Use include_user_info_headers() helper instead of manual header construction
- Rename json_data to form_data for consistency
- Rename elem to item in response parsing
- Remove unnecessary retry loop from Azure function
- Restore lambda wrappers with user parameter to preserve user context
- Revert reranking function signature to match dev branch
- Restore separate condition checks with warning logs

External APIs (OpenAI, Azure, Ollama) already use async aiohttp - unaffected.