llama.cpp

server: use pos_next instead of n_tokens for m-rope

#22439

Merged