Instead of creating your own engine, would it really be that hard to add paged attention to llama.cpp?