Add LiteLLM integration for multi-provider model support

## Proposal

Add LiteLLM integration to support multiple vision-capable model providers instead of only Anthropic.

### Design Document

See [litellm-brainstorming.md](https://github.com/hamelsmu/ai-jup/blob/main/litellm-brainstorming.md) for full details.

### Summary

**Allowed Providers:**
- OpenAI
- Anthropic
- Gemini

**Key Features:**
1. **Model Picker UI** - Provider dropdown → Model dropdown (filtered, sorted by release date)
2. **Vision-only models** - Use `litellm.supports_vision()` to filter
3. **Test Connection button** - Verify API keys work before execution
4. **Global settings** - Model selection in JupyterLab settings, not per-cell

**Why LiteLLM:**
- Unified API across providers
- Built-in vision support detection
- Handles message format translation (images, tool calls)
- Actively maintained model database

### Implementation Phases

1. Backend LiteLLM integration + `/ai-jup/models` endpoint
2. Frontend model picker component
3. Test connection endpoint + validation

Start by reading litellm-brainstorming.md, then map the three phases: the backend /ai-jup/models endpoint, the frontend model picker, and the test connection endpoint with validation. Done means supported vision-capable OpenAI, Anthropic, and Gemini models can be selected in JupyterLab settings and connections are verified before execution.

Written by the indexing model from the issue text.