### Check for existing issues

- [x] I have searched the existing issues and checked that my issue is not a duplicate.

### The Feature

## Summary

When using Gemini image models through the official Google GenAI SDK / API, response parts can include `part.thought`. This allows users to distinguish interim "thought images" from final generated images.

However, when using Gemini image models through LiteLLM, generated images are normalized into `response.choices[0].message.images`. As far as I can tell, the original Gemini `part.thought` metadata is not exposed in the normalized LiteLLM response. As a result, users cannot reliably tell whether an image is a thought image or the final output image.

I would like LiteLLM to preserve this metadata when normalizing Gemini image responses.

## Why this matters

Gemini image models can return intermediate thinking outputs as part of their generation process. In the official SDK, these can be identified with `part.thought`.

This distinction is important for applications that need to:

* filter out thought images * select only the final generated image * display thought images separately for debugging or transparency * avoid accidentally…