v3.0.0

-Move from a "request only" protocol to a "request/response". This makes latency measurements more meaningful and consistent between TX and RX requests. With the "request only" protocol we'd stop timing when the final bytes of the payload were buffered on the client's socket, which would exclude the time it took those bytes to actually get to the service.

-Set TCP_NODELAY on sockets unconditionally. For small requests, without this option we would be blocked by Nagle due to sending the small request message followed immediately by the small payload (or likewise on the service side due to sending the small payload followed by the small response message). I think there's no case for Nagle in this tool.