vision: add MessagePart::Image and supports_vision to LlmProvider trait

[bug-ops]

Parent: #490

Scope

Foundation layer — extend the content model and provider trait for image support.

Changes

1. crates/zeph-llm/src/provider.rs

   • Add MessagePart::Image { data: Vec<u8>, mime_type: String } variant to the enum
   • Add fn supports_vision(&self) -> bool with default false to LlmProvider trait
   • Update flatten_parts() to skip image parts (text-only fallback)

2. crates/zeph-core/src/channel.rs

   • Add ImageData { data: Vec<u8>, mime_type: String, filename: Option<String> } struct
   • Add images: Vec<ImageData> field to ChannelMessage

3. Agent loop (crates/zeph-core/src/agent.rs or equivalent)

   • Convert ChannelMessage.images into MessagePart::Image when building LLM messages

Notes

• No provider implementations yet — just the trait and data model
• This unblocks all downstream issues