google-deepmind / tips Jupyter Notebook 647 41 TIPSv2 (CVPR'26) and TIPS (ICLR'25) image-text spatial-understanding vision-encoder