feat: add db-backed OpenAI-compatible model management

[yjing86]

Summary

This PR now includes several commits on the same branch, with the main change being the introduction of database-backed OpenAI-compatible model management for Sugar-AI.

Main change

• replace embedded Hugging Face runtime inference with an OpenAI-compatible provider layer
• persist model configuration in the database
• support runtime activation and switching of the active model
• add admin APIs and UI for model management
• update Docker, env examples, and README for Ollama/vLLM or external OpenAI-compatible backends
• add an API smoke test script

Additional branch commits

This PR branch also currently includes a few smaller commits that landed before the main feature commit:

• refine the GSoC 2026 AI optimization proposal
• restore Jinja TemplateResponse compatibility
• fix loading DEV_MODE from environment settings

Why

Sugar-AI was previously tightly coupled to local in-process model loading. This change moves the app toward a provider-oriented architecture and makes model configuration and switching easier to operate.

Testing

• ./.venv/bin/python scripts/test_api.py --internal

Notes

• this PR currently contains 4 commits because the branch includes earlier fixes and docs work
• the main functional change is the OpenAI-compatible model management flow
• Resolves support user-provided AI model providers (OpenAI, Anthropic, Ollama) #51
• Resolves refactor: decouple model inference from the application and support pluggable inference backends #117