[SPARK-8392] RDDOperationGraph: getting cached nodes is slow

```def getAllNodes: Seq[RDDOperationNode] =
{ _childNodes ++ _childClusters.flatMap(_.childNodes) }```

when the ```_childClusters``` has so many nodes, the process will hang on. I think we can improve the efficiency here.

Author: xutingjun <xutingjun@huawei.com>

Closes #6839 from XuTingjun/DAGImprove and squashes the following commits:

53b03ea [xutingjun] change code to more concise and easier to read
f98728b [xutingjun] fix words: node -> nodes
f87c663 [xutingjun] put the filter inside
81f9fd2 [xutingjun] put the filter inside