Fix tx-based latency calculations

Summary:
We need to take start timestamp off of the very first tx byte, not the
last one

Reviewed By: jbeshay

Differential Revision: D66476343

fbshipit-source-id: de55014a103f0dada3c363a3c340fab38e8ac633

quic/tools/tperf/tperf.cpp

@@ -413,7 +413,7 @@ class ServerStreamHandler : public quic::QuicSocket::ConnectionSetupCallback,
     auto sendBuffer = buf_->clone();
     sendBuffer->append(blockSize_);
     CHECK_GT(blockSize_, 0);
-    auto r = sock_->registerTxCallback(*stream, blockSize_ - 1, this);
+    auto r = sock_->registerTxCallback(*stream, 0, this);
     if (r.hasError()) {
       LOG(FATAL) << "Got error on registerTxCallback: "
                  << quic::toString(r.error());