Add task queue/runtime stats per LLAP daemon to output

There are several scenarios, where we want to investigate if a particular LLAP daemon is performing faster or slower than the others in the cluster. In these scenarios, it is specifically important to figure out if tasks spent significant time, waiting for an available executor (queued) vs. on the execution itself. Also, a skew in task-to-daemon assignment is interesting.

This patch adds these statistics to the TezCounters and therefore to the job output on a per LLAP daemon base. Here is an example.

INFO : LlapTaskRuntimeAgg by daemon:
INFO :    Count-host-1.example.com: 41
INFO :    Count-host-2.example.com: 39
INFO :    Count-host-3.example.com: 45
INFO :    QueueTime-host-1.example.com: 51437776
INFO :    QueueTime-host-2.example.com: 35758306
INFO :    QueueTime-host-3.example.com: 47168327
INFO :    RunTime-host-1.example.com: 165151539295
INFO :    RunTime-host-2.example.com: 141729193528
INFO :    RunTime-host-3.example.com: 166876988771

The "Count-" are simple task counts for the appended host name (LLAP daemon)

The "QueueTime-" values tell, how long tasks waited in the TaskExecutorService's queue before getting actually executed.

The "RunTime-" values cover the time from execution start to finish (where finish can either be successful execution or a killed/failed execution).

For the new counts to appear in the output, both - the preexisting hive.tez.exec.print.summary and the new hive.llap.task.time.print.summary have to be set to true.

 
<property>
  <name>hive.tez.exec.print.summary</name>
  <value>true</value>
</property>
<property>
  <name>hive.llap.task.time.print.summary</name>
  <value>true</value>
</property>