vLLM adds encoder timing statistics to Model Runner V2

The vLLM project has implemented support for encoder timing statistics within its Model Runner V2 architecture. This enables more granular performance profiling during the inference process.