Fix inference loop logic and improve image resizing

[RubyNg]

Summary

This PR addresses two critical issues in inference_demo.py:

1. Conversation Logic Fix: The original code failed to append the model's response to history and used the wrong role for execution results, causing infinite loops.
2. Visual Grounding Fix: Implemented smart_resize strategy (referenced from VLMEvalKit) to fix inaccurate bounding box predictions caused by image downscaling.

Changes

• Appended response_message to chat_message to maintain context.
• Changed execution result role from assistant to user.
• Implemented smart_resize to preserve image details for better coordinate grounding.

Verified

Tested locally with Visual Probe tasks. The model now correctly generates answers and crops accurate regions.