[SPARK-5885][MLLIB] Add VectorAssembler as a feature transformer

[mengxr]

VectorAssembler merges multiple columns into a vector column. This PR contains content from #5195.

carry ML attributes (moved to a follow-up PR)

[SparkQA]

Test build #29199 has finished for PR 5196 at commit 29fb6ac.

• This patch fails Spark unit tests.
• This patch merges cleanly.
• This patch adds the following public classes (experimental):
  • class VectorAssembler extends Transformer with HasInputCols with HasOutputCol
  • case class CreateStruct(children: Seq[Expression]) extends Expression

[liancheng]

The error message is confusing. CreateStruct can be called with unresolved children (and you actually did this here), but CreateStruct.dataType can't, because child types may be unknown.

[liancheng]

The build failure was caused by codegen. When the ExpressionEvaluationSuite.CreateStruct test case is executed separately, codegen is not enabled by default, and CreateStruct.dataType is not touched. But on Jenkins, some test suite executed earlier enabled codegen, then CreateStruct.dataType got called and threw exception because BoundReference used in the test case if not NamedExpression.

[liancheng]

Actually, the failed test suites are GeneratedEvaluationSuite and GeneratedMutableEvaluationSuite, both of which just extend ExpressionEvaluationSuite.

[petro-rudenko]

Maybe call it FeatureUnion to keep the same semantic with sklearn

[oefirouz]

+1. I like the idea of trying to keep naming and architecture as similar to sklearn as possible.

[mengxr]

This is not exactly sklearn's FeatureUnion that takes transformers. VectorAssembler merges arbitrary columns. I'm a little worried that if we use the same name people would expect the same behavior.

[SparkQA]

Test build #30018 has finished for PR 5196 at commit a52b101.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds the following public classes (experimental):
  • class VectorAssembler extends Transformer with HasInputCols with HasOutputCol
  • class SparseMatrix(Matrix):
• This patch does not change any dependencies.

[mengxr]

Merged into master. Will send a follow-up PR to handle attributes.