[SPARK-25051][SQL] FixNullability should not stop on AnalysisBarrier

## What changes were proposed in this pull request?

The introduction of `AnalysisBarrier` prevented `FixNullability` to go through all the nodes. This introduced a bug, which can lead to wrong results, as the nullability of the output attributes of an outer join can be wrong.

The PR makes `FixNullability` going through the `AnalysisBarrier`s.

## How was this patch tested?

added UT

Author: Marco Gaido <marcogaido91@gmail.com>

Closes #22102 from mgaido91/SPARK-25051.

sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/analysis/Analyzer.scala

@@ -1704,6 +1704,7 @@ class Analyzer(
 
     def apply(plan: LogicalPlan): LogicalPlan = plan transformUp {
       case p if !p.resolved => p // Skip unresolved nodes.
+      case ab: AnalysisBarrier => apply(ab.child)
       case p: LogicalPlan if p.resolved =>
         val childrenOutput = p.children.flatMap(c => c.output).groupBy(_.exprId).flatMap {
           case (exprId, attributes) =>

sql/core/src/test/scala/org/apache/spark/sql/DataFrameSuite.scala

@@ -2300,4 +2300,10 @@ class DataFrameSuite extends QueryTest with SharedSQLContext {
       checkAnswer(aggPlusFilter1, aggPlusFilter2.collect())
     }
   }
+
+  test("SPARK-25051: fix nullabilities of outer join attributes doesn't stop on AnalysisBarrier") {
+    val df1 = spark.range(4).selectExpr("id", "cast(id as string) as name")
+    val df2 = spark.range(3).selectExpr("id")
+    assert(df1.join(df2, Seq("id"), "left_outer").where(df2("id").isNull).collect().length == 1)
+  }
 }