Model card: multimodal embeddings now run on GPU (LAST pooling end-state); re-ingest note

Model card: correct multimodal-embedding framing (Gemma 4 12B native interleave; dedicated graph is the stopgap)

Model card: clarify GPU multimodal behavior (media indexes via text bridge; raw vectors auto-skipped on GPU)

Model card: GPU usage section + hardware-verified note (2x RTX 4090, CUDA 12.8)

Upload README.md with huggingface_hub