Fetch request metrics are updated twice for delayed fetches

We track fetch request rate metrics in `BrokerTopicStats`. We have both topic level and total metrics. Currently these values are updated in `ReplicaManager.readFromLocalLog`. Inside `fetchMessages`, we invoke `readFromLocalLog` first to see if the fetch can be immediately satisfied. If not, then we put it in purgatory. After completing purgatory, we make a second call to `readFromLocalLog` which will again update the request rate metrics. Hence we are overcounting the fetch request rate.