Add local Ollama/Qwen support through the OpenAI-compatible provider

From the PR description

Summary

• add OpenAI-compatible local runtime support for self-hosted endpoints like Ollama
• add Qwen model options in Mike's model picker and document the Docker-plus-native-Ollama setup
• make the OpenAI-compatible path stream local model output incrementally so Mike surfaces answers instead of appearing blank while Ollama is still generating

What changed

• made the OpenAI adapter configurable via OPENAI_COMPAT_BASE_URL, OPENAI_COMPAT_MODEL, and OPENAI_COMPAT_ENDPOINT_MODE
• added support for chat-completions style OpenAI-compatible backends used by Ollama and vLLM
• treated env-configured OpenAI-compatible runtimes as available even without a hosted OpenAI API key
• added Qwen 3 8B and 14B options to the model picker
• documented the local Docker app stack plus native Ollama workflow for testing Mike locally

Validation

• npm run build --prefix backend
• local Docker stack rebuilt with the Ollama-compatible env path
• verified local Supabase keys were wired into the Docker env files for the local dev stack

Notes

• frontend lint still reports existing repo-wide issues unrelated to this branch
• this branch is intentionally scoped to local inference enablement first; broader benchmark work can build on top

Screenshot