optimizations

llama.cpp CPU Offload Optimizations

Explore CPU offload optimizations for llama.cpp: disable CUDA graphs and target FFN sub-layers.