llama.cpp

Improve utilization (and tg t/s) by reducing K_QUANTS_PER_ITERATION to 1 on DMMV path

#25063

Merged