Scale LLM serving with programmable cross-engine serving patterns, all in a few lines of Python