vLLM serving stack ported to C++20 with 66 MiB binary

Community port of vLLM's serving stack compiles to a 66 MiB C++20 binary with no Python at inference, verified token-for-token against vLLM. Author built it to avoid vLLM's 9.1 GiB virtualenv install and embed inference; project is unaffiliated and not endorsed by vLLM.