Card: decode speed after the attention change

Decoder: concatenate attention scores instead of K/V (no per-layer cache copy, ~17% faster decode, same output)

Unlimited-OCR Core ML: fp16 page encoder + int8 stateful multi-page decoder

initial commit