CLIP ViT-B/32 image encoder (litert-torch, GPU-clean 4D attention) + text embeddings

initial commit