Local Ollama backend for the AI bar

The AI bar (`Ctrl+K`) currently hits OpenAI or Anthropic. A local **Ollama** option would let users run NL→SQL fully offline.

**Where**: \`src/ai/client.rs\` (add \`Backend::Ollama\`), \`src/ai/config.rs\` (config schema), \`src/ai/prompt.rs\` (may need to trim — local models have smaller context windows)

**Acceptance**
- \`xeli config set-key ollama http://localhost:11434\` works
- \`Ctrl+K\` → query goes to \`POST /api/generate\` with the configured model (default \`qwen2.5-coder:7b\` or similar)
- Falls back to clear error if the daemon isn't running

Bonus: streaming response into the bar.