[SPARK-13096][TEST] Fix flaky verifyPeakExecutionMemorySet

Previously we would assert things before all events are guaranteed to have been processed. To fix this, just block until all events are actually processed, i.e. until the listener queue is empty.

https://amplab.cs.berkeley.edu/jenkins/job/spark-master-test-sbt-hadoop-2.7/79/testReport/junit/org.apache.spark.util.collection/ExternalAppendOnlyMapSuite/spilling/

Author: Andrew Or <andrew@databricks.com>

Closes #10990 from andrewor14/accum-suite-less-flaky.

core/src/test/scala/org/apache/spark/AccumulatorSuite.scala

@@ -307,6 +307,8 @@ private[spark] object AccumulatorSuite {
     val listener = new SaveInfoListener
     sc.addSparkListener(listener)
     testBody
+    // wait until all events have been processed before proceeding to assert things
+    sc.listenerBus.waitUntilEmpty(10 * 1000)
     val accums = listener.getCompletedStageInfos.flatMap(_.accumulables.values)
     val isSet = accums.exists { a =>
       a.name == Some(PEAK_EXECUTION_MEMORY) && a.value.exists(_.asInstanceOf[Long] > 0L)