elitan adds Ollama as a fourth LLM provider for fully local inference

A new 227-line backend/src/lib/llm/ollama.ts speaks directly to Ollama's streaming chat endpoint, defaulting to http://localhost:11434 with an override via OLLAMA_BASE_URL. Models prefixed with ollama: are routed through it; the prefix is stripped before the request goes on the wire. Streaming, tool calls, and text completion all go through the same StreamChatResult contract the other providers use, so the rest of the backend sees no difference.

OLLAMA_ENABLED=false in .env.example means the feature is opt-in. Frontend changes in ModelToggle.tsx and modelAvailability.ts mark Ollama models as always-available, removing the API-key-missing warning that appears for cloud providers. The account/models page is updated to match. Pre-registered models are ollama:llama3.1 and ollama:qwen3:8b.

Two things worth reviewing before relying on tool use: the streaming JSON parser in ollama.ts is hand-rolled rather than using a library, and tool-call argument coercion accepts both object and JSON-string shapes permissively. That's fine for general chat but could produce odd results with small local models that don't format tool responses reliably.

This commit sits alongside Supabase removal and docker-compose work in the same day's run - elitan is clearly driving toward a self-hostable, zero-external-dependency deployment.