v4: two-lane decoder (two pieces per infer call, grouped-query attention)

Confucius4-R2T2 converted for Core ML on the Apple Neural Engine (encoder fp16, decoder LUT8 with 128-row prefill and verify head)