Add paper, project, and code links; fix task category to image-to-video

Ship the val split's own pose ground truth

Add metadata.csv (benchmark prompts): the documented eval path needs it

Add rgb + reference_image for the 600 benchmark clips: the benchmark is now self-contained

Add dwpose control signals (DWPose is ours, not shipped with OpenHumanVid)

Add control signals (depth/scribble), pose GT keypoints, and split file lists

initial commit