vision: Ollama provider image support with configurable vision model

[bug-ops]

Parent: #490
Depends on: #491

Scope

Implement vision support in the Ollama provider, including a separate configurable model for image processing.

Changes

1. crates/zeph-core/src/config/types.rs

   • Add vision_model: Option<String> field to Ollama config section

2. config/default.toml

   • Add commented # vision_model = "llava" example

3. crates/zeph-llm/src/ollama.rs

   • Override supports_vision() -> true when vision_model is configured
   • In convert_message(): extract MessagePart::Image data, pass via ChatMessage::with_images()
   • When vision_model is set and message contains images: temporarily switch to vision model for that request, then switch back

Notes

• ollama-rs ChatMessage already has .with_images() — just needs wiring
• Vision models (llava, bakllava, moondream) are pulled separately in Ollama
• Test with mock: verify images are passed through to the API request