Exported and not published. On an iPhone 16 MLX decodes at 3.2 ms/step (int8) and
3.9 ms/step (bf16) against 1.95 ms for XNNPACK int8, at the same BLEU. Use `xnnpack/`.