Support token metrics

[luohua13]

VLLM runtime has a wealth of token metrics, example prompt_tokens_total and generation_tokens_total. Why does mlserver have none?

[luohua13]

I add two token metrics model_input_tokens and model_output_tokens in pr #2006