Automatic model routing: route image inputs to vision-capable model, return text to text-only model

### Describe the feature or problem you''d like to solve

When using a text-only model like DeepSeek as the primary model, image inputs are either ignored or cause errors, because DeepSeek lacks vision capabilities. Users must manually switch models with /model every time they want to include an image, which breaks workflow flow.

### Proposed solution

Automatic model routing based on input modality. When a user''s prompt includes images (pasted, drag-and-drop, or @-referenced), Copilot CLI should:

1. Detect that the primary model doesn''t support vision
2. Auto-route the image(s) to a vision-capable model (e.g., GPT-4o, Claude Sonnet 4.5) configured as the "vision fallback"
3. Receive a text description of the image(s) from the vision model
4. Inject that text description into the prompt sent to the primary (text-only) model
5. The user only sees the final response from their chosen primary model, with the image description transparently provided as context

Configuration (example .copilot/config or copilot-instructions.md):
```
vision_fallback_model: "gpt-4o"
vision_fallback_behavior: "describe_and_forward"
```

### Example prompts or workflows

1. User has DeepSeek selected. Pastes a screenshot of a UI bug and types "Fix this layout issue". CLI detects image + text-only model -> sends image to GPT-4o -> GPT-4o returns text description -> that description is prepended to DeepSeek''s prompt -> DeepSeek fixes the code.

2. User @-references a diagram.png file: "Implement this architecture diagram". Same routing flow - architecture described in text, DeepSeek implements.

3. /model deepseek is active. User pastes error screenshot. Without needing to switch models, user gets a code fix.

4. Works in reverse too - if the primary model IS vision-capable, no routing occurs; image is sent directly.

5. Configurable: user can disable auto-routing or choose which vision model to use as fallback.

### Additional context

- This is inspired by the existing sub-agent and delegation architecture already present in Copilot CLI
- Reduces friction for users who prefer text-only models for cost/performance but occasionally need vision capabilities
- Could be implemented as a lightweight pre-processing step before the main model invocation
- Related: the /fleet and custom agents infrastructure could potentially be leveraged for this

Start by reading the existing sub-agent and delegation architecture, then inspect the /fleet and custom agents infrastructure mentioned in the issue. Define how image detection, vision fallback configuration, description forwarding, and the no-routing path should fit together. Done means image inputs reach a configured vision-capable fallback when needed, while users receive only the primary model’s final response.

Escrita pelo modelo de indexação a partir do texto da issue.