Update with arXiv link

Clarify that input is LeJEPA embeddings, not raw CT images; LeJEPA is a required dependency

Fix model tree: base_model is Llama 3.2 3B only, LeJEPA provides input embeddings not lineage

Remove optimized threshold results from model card

Rewrite model card: describe Ker-VLJEPA-3B as multi-modal VLM, not embedding model

Initial release: Ker-VLJEPA-3B inference package

initial commit