llama.cpp

Optimization: Qwen3 next autoregressive pass

#17996

Merged