[YARN][DOC] Increasing NodeManager's heap size with External Shuffle Service

## What changes were proposed in this pull request?

Suggest users to increase `NodeManager's` heap size if `External Shuffle Service` is enabled as
`NM` can spend a lot of time doing GC resulting in  shuffle operations being a bottleneck due to `Shuffle Read blocked time` bumped up.
Also because of GC  `NodeManager` can use an enormous amount of CPU and cluster performance will suffer.
I have seen NodeManager using 5-13G RAM and up to 2700% CPU with `spark_shuffle` service on.

## How was this patch tested?

#### Added step 5:
![shuffle_service](https://cloud.githubusercontent.com/assets/15244468/20355499/2fec0fde-ac2a-11e6-8f8b-1c80daf71be1.png

)

Author: Artur Sukhenko <artur.sukhenko@gmail.com>

Closes #15906 from Devian-ua/nmHeapSize.

(cherry picked from commit 55589987)
Signed-off-by: Reynold Xin <rxin@databricks.com>

docs/running-on-yarn.md

@@ -559,6 +559,8 @@ pre-packaged distribution.
 1. In the `yarn-site.xml` on each node, add `spark_shuffle` to `yarn.nodemanager.aux-services`,
 then set `yarn.nodemanager.aux-services.spark_shuffle.class` to
 `org.apache.spark.network.yarn.YarnShuffleService`.
+1. Increase `NodeManager's` heap size by setting `YARN_HEAPSIZE` (1000 by default) in `etc/hadoop/yarn-env.sh` 
+to avoid garbage collection issues during shuffle. 
 1. Restart all `NodeManager`s in your cluster.
 
 The following extra configuration options are available when the shuffle service is running on YARN: