llama : add --n-cpu-ffn option by John-194 · Pull Request #26622 · ggml-org/llama.cpp

option similar to the existing --n-cpu-moe It puts user specified amount of FFN sublayers for dense models. tl;dr faster dense models for low VRAM people