add fp16 vision (default), ANE-layout fp16 vision, fp16 heads for all buckets; manifest variants + compute-unit guidance

card: Swift runtime parity and latency

add vision pos_embed table, head buckets L512/L2048, manifest text_config_summary

clef-vision-0.8b: Core ML packages (vision fp32, LM L512/L1024/L2048 fp16, head fp32), embeddings, tokenizer, card

initial commit