MCP image content blocks stripped from tool responses — AI cannot view images

### Description

When an MCP tool returns image content blocks (e.g., base64-encoded JPEG/PNG from `photoshop_get_preview`), the Copilot client strips or truncates the image data before passing the response to the AI model. The AI receives only the text/metadata portion and has no way to access the actual image.

### Expected Behavior

Image content blocks returned by MCP tools should be preserved and forwarded to the AI model, allowing it to "see" the image and provide visual-aware assistance.

### Actual Behavior

1. The MCP tool returns image data correctly (e.g., `photoshop_get_preview` returns `bytes: 30115`, indicating actual image content was generated)
2. The Copilot client strips the image content block from the response
3. The AI only receives text metadata: `{"ok": true, "width": 335, "height": 333, "bytes": 30115}`
4. The AI has no way to view the image unless the user manually takes a screenshot and sends it as a chat attachment

### Impact

- AI cannot verify visual editing results autonomously
- Cannot provide intelligent suggestions based on image content
- Requires manual screenshot sharing for every visual feedback loop, significantly reducing efficiency

### Reproduction Steps

1. Connect to a Photoshop MCP server, for example https://github.com/alisaitteke/photoshop-mcp
2. Open a document in Photoshop
3. Call `photoshop_get_preview`
4. Observe that the AI's response contains no image data

Reproduce the issue with a Photoshop MCP server and the photoshop_get_preview tool, then trace where the Copilot client handles the MCP tool response before it reaches the AI model. Done means image content blocks are preserved and forwarded so the model can access the image as well as the text metadata.

索引モデルが issue の本文から書いたものです。