Moving Kafka request latency measurement to the server side

[Xi Yang]

Clients in the Kafka benchmark send requests to the server asynchronously where requests can be buffered in the queue before being processed by the server. Now the benchmark measures the latency of request from the time of sending the requests to the time of receiving the response from clients, which adds the queueing delay to the request latency.

To avoid this issue, we should move the measurement to the server side starting from receiving the request in the server to the time of sending out the response.