[SPARK-14322][MLLIB] Use treeAggregate instead of reduce in OnlineLDAOptimizer

## What changes were proposed in this pull request?
jira: https://issues.apache.org/jira/browse/SPARK-14322

OnlineLDAOptimizer uses RDD.reduce in two places where it could use treeAggregate. This can cause scalability issues. This should be an easy fix.
This is also a bug since it modifies the first argument to reduce, so we should use aggregate or treeAggregate.
See this line: https://github.com/apache/spark/blob/f12f11e578169b47e3f8b18b299948c0670ba585/mllib/src/main/scala/org/apache/spark/mllib/clustering/LDAOptimizer.scala#L452
and a few lines below it.

## How was this patch tested?
unit tests

Author: Yuhao Yang <hhbyyh@gmail.com>

Closes #12106 from hhbyyh/ldaTreeReduce.

(cherry picked from commit 8cffcb6)
Signed-off-by: Joseph K. Bradley <joseph@databricks.com>

Author: hhbyyh

mllib/src/main/scala/org/apache/spark/mllib/clustering/LDAOptimizer.scala

@@ -449,10 +449,11 @@ final class OnlineLDAOptimizer extends LDAOptimizer {
       }
       Iterator((stat, gammaPart))
     }
-    val statsSum: BDM[Double] = stats.map(_._1).reduce(_ += _)
+    val statsSum: BDM[Double] = stats.map(_._1).treeAggregate(BDM.zeros[Double](k, vocabSize))(
+      _ += _, _ += _)
     expElogbetaBc.unpersist()
     val gammat: BDM[Double] = breeze.linalg.DenseMatrix.vertcat(
-      stats.map(_._2).reduce(_ ++ _).map(_.toDenseMatrix): _*)
+      stats.map(_._2).flatMap(list => list).collect().map(_.toDenseMatrix): _*)
     val batchResult = statsSum :* expElogbeta.t
 
     // Note that this is an optimization to avoid batch.count