GPU observability for LLM inference: the twelve DCGM and vLLM metrics that dictate the health of your productionJun 1, 2026