[SPARK-20688][SQL] correctly check analysis for scalar sub-queries

[cloud-fan]

What changes were proposed in this pull request?

In CheckAnalysis, we should call checkAnalysis for ScalarSubquery at the beginning, as later we will call plan.output which is invalid if plan is not resolved.

How was this patch tested?

new regression test

[cloud-fan]

cc @rxin @gatorsmile

[SparkQA]

Test build #76725 has started for PR 17930 at commit 3ccff2e.

[gatorsmile]

LGTM

[gatorsmile]

retest this please

[wzhfy]

Not introduced by this pr, but I just saw a if-else style nit here

             if (conditions.isEmpty && query.output.size != 1) {
                failAnalysis(
                 s"Scalar subquery must return only one column, but got ${query.output.size}")
             }
             else if (conditions.nonEmpty) {

[SparkQA]

Test build #76733 has finished for PR 17930 at commit 3ccff2e.

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

[SparkQA]

Test build #76739 has finished for PR 17930 at commit 013137f.

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

[cloud-fan]

thanks for the review, merging to master/2.2/2.1!