Use millisecond granularity for publisher ramp-up period

It was seconds, which works for very long ramp-up period
(e.g. 60 seconds or more) and a small number of publishers
(few 100s), but not short ramp-up period and large numbers
of publishers (start 5000 publishers in 15 seconds). In such
cases, the workload was not random enough like in real-life
and metrics like publisher confirm latency could be much
higher than expected.

References #394

src/main/java/com/rabbitmq/perf/MulticastSet.java

@@ -456,8 +456,9 @@ PRODUCER_THREAD_PREFIX, nbThreadsForProducerScheduledExecutorService(params)
             Supplier<Duration> startDelaySupplier;
             if (params.getProducerRandomStartDelayInSeconds() > 0) {
                 Random random = new Random();
-                startDelaySupplier = () -> Duration.ofSeconds(
-                    random.nextInt(params.getProducerRandomStartDelayInSeconds()) + 1
+                int bound = params.getProducerRandomStartDelayInSeconds() * 1000;
+                startDelaySupplier = () -> Duration.ofMillis(
+                    random.nextInt(bound) + 1
                 );
             } else {
                 startDelaySupplier = () -> Duration.ZERO;