Add FetchMetricLatency metric

[liu-cong]

Low metric fetch latency is crucial as the load balancing algorithm relies on near real time metrics from the model servers. Therefore it's important to monitor the metric fetch latency.
Currently we simply log the latency, we should integrate Prometheus and add a Histogram metric.