[Ml] SPARK-5804 Explicitly manage cache in Crossvalidator k-fold loop

On a big dataset explicitly unpersist train and validation folds allows to load more data into memory in the next loop iteration. On my environment (single node 8Gb worker RAM, 2 GB dataset file, 3 folds for cross validation), saved more than 5 minutes.

Author: Peter Rudenko <petro.rudenko@gmail.com>

Closes #4595 from petro-rudenko/patch-2 and squashes the following commits:

66a7cfb [Peter Rudenko] Move validationDataset cache to declaration
c5f3265 [Peter Rudenko] [Ml] SPARK-5804 Explicitly manage cache in Crossvalidator k-fold loop

mllib/src/main/scala/org/apache/spark/ml/tuning/CrossValidator.scala

@@ -108,6 +108,7 @@ class CrossValidator extends Estimator[CrossValidatorModel] with CrossValidatorP
       // multi-model training
       logDebug(s"Train split $splitIndex with multiple sets of parameters.")
       val models = est.fit(trainingDataset, epm).asInstanceOf[Seq[Model[_]]]
+      trainingDataset.unpersist()
       var i = 0
       while (i < numModels) {
         val metric = eval.evaluate(models(i).transform(validationDataset, epm(i)), map)
@@ -115,6 +116,7 @@ class CrossValidator extends Estimator[CrossValidatorModel] with CrossValidatorP
         metrics(i) += metric
         i += 1
       }
+      validationDataset.unpersist()
     }
     f2jBLAS.dscal(numModels, 1.0 / map(numFolds), metrics, 1)
     logInfo(s"Average cross-validation metrics: ${metrics.toSeq}")