[DSV3] GroupedExperts weights conversion optimization (#1639)

## Algorithm summary:

<img width="851" height="589" alt="Screenshot 2025-08-28 at 3 27 36 PM"
src="https://github.com/user-attachments/assets/87c38d6f-7ed0-47dc-a599-d83ddee9c3b8"
/>

## Numerical comparison using 16B model:
- Note: the final results are not exactly same because the numerics are
small

<img width="992" height="676" alt="Screenshot 2025-08-28 at 3 25 11 PM"
src="https://github.com/user-attachments/assets/a0af55a3-2cdb-4108-82d8-82be246f1ff5"
/>
<img width="992" height="662" alt="Screenshot 2025-08-28 at 3 26 41 PM"
src="https://github.com/user-attachments/assets/57009581-db6f-4ba3-b1ad-0a95f80bbe67"
/>

Author: wwwjn