Support local LLM servers (Ollama / llama.cpp / LM Studio)

[8crsk]

Users running Ollama or llama.cpp on a home machine should be able to point 4AIs at it — the ultimate no-cloud setup: phone agent + local model, zero external API calls.

Design questions to settle first (comment here!)

• Custom base-URL field in Settings vs a dedicated 'Local server' provider entry?
• How to handle servers with no API key (the gateway config currently always sends one)?
• mDNS/manual discovery of the server on the LAN?

Where: ProviderCatalog.kt + the Settings provider UI (ui/settings/). The gateway already speaks openai-completions to any baseUrl.

Acceptance criteria

• User can enter an arbitrary OpenAI-compatible baseUrl (+ optional key)
• Works with Ollama's /v1 endpoint end-to-end
• Graceful error surface when the server is unreachable