Add images: EmbeddingGemma2Vision (vision_70/140/280, GPU fp16) + position tables

EmbeddingGemma 2 text encoder on Core ML (Neural Engine): 7 shared-weight functions, bf16 token table

initial commit