What is Copy-on-Write (CoW) optimization at the operating system level? How did changing the reference counting logic for immortal objects (Immortal Objects, PEP 683) help reduce memory consumption when deploying web applications via multiprocessing, such as Gunicorn with the --preload flag?