[SPARK-11803][SQL] fix Dataset self-join

[cloud-fan]

When we resolve the join operator, we may change the output of right side if self-join is detected. So in Dataset.joinWith, we should resolve the join operator first, and then get the left output and right output from it, instead of using left.output and right.output directly.

[cloud-fan]

cc @marmbrus @rxin @gatorsmile

[SparkQA]

Test build #46210 has finished for PR 9806 at commit 2c9f485.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[gatorsmile]

Your fix looks pretty clean to me. Let me share my test cases this PR failed.

 test("joinWith tuple - self join 1") {
   val ds = Seq(("a", 1), ("b", 2)).toDS()
   ds.joinWith(ds, $"_2" === $"_2").collect()
 }

  test("joinWith tuple - self join 2") {
    val ds1 = Seq(("a", 1), ("b", 2)).toDS()
    val ds2 = Seq(("a", 1), ("b", 2)).toDS().as("a")
    ds1.joinWith(ds2, $"_2" === $"a._2").collect()
  }

Do you want me to send a separate PR based on your changes if you are busy? Or you will fix them in this PR? Either is fine to me. : )

Thank you!

[marmbrus]

LGTM, merging to master and 1.6.

@gatorsmile please open JIRAs targeted at 1.6.0 for the bugs you have found. (also use checkAnswer when writing test cases). Thanks!

[gatorsmile]

Sure. Will do. Thanks!