timm adds multi-label image classification support

A timm contributor says they added multi-label classification support and uploaded datasets with a consistent schema to Hugging Face. In their tests, a timm fine-tune and linear probe performed much better and faster than the VLM approaches they tried using the same ViT encoders. They also say they used Claude and Codex to curate long-tail labels from open-access Met Museum data for an iMet-like dataset; licensing constraints prevented them from uploading the original Kaggle version.