[SPARK-6037][SQL] Avoiding duplicate Parquet schema merging

`FilteringParquetRowInputFormat` manually merges Parquet schemas before computing splits. However, it is duplicate because the schemas are already merged in `ParquetRelation2`. We don't need to re-merge them at `InputFormat`.

Author: Liang-Chi Hsieh <viirya@gmail.com>

Closes #4786 from viirya/dup_parquet_schemas_merge and squashes the following commits:

ef78a5a [Liang-Chi Hsieh] Avoiding duplicate Parquet schema merging.

(cherry picked from commit 4ad5153)
Signed-off-by: Cheng Lian <lian@databricks.com>

sql/core/src/main/scala/org/apache/spark/sql/parquet/ParquetTableOperations.scala

@@ -434,22 +434,13 @@ private[parquet] class FilteringParquetRowInputFormat
      return splits
     }
 
-    Option(globalMetaData.getKeyValueMetaData.get(RowReadSupport.SPARK_METADATA_KEY)).foreach {
-      schemas =>
-        val mergedSchema = schemas
-          .map(DataType.fromJson(_).asInstanceOf[StructType])
-          .reduce(_ merge _)
-          .json
-
-        val mergedMetadata = globalMetaData
-          .getKeyValueMetaData
-          .updated(RowReadSupport.SPARK_METADATA_KEY, setAsJavaSet(Set(mergedSchema)))
-
-        globalMetaData = new GlobalMetaData(
-          globalMetaData.getSchema,
-          mergedMetadata,
-          globalMetaData.getCreatedBy)
-    }
+    val metadata = configuration.get(RowWriteSupport.SPARK_ROW_SCHEMA)
+    val mergedMetadata = globalMetaData
+      .getKeyValueMetaData
+      .updated(RowReadSupport.SPARK_METADATA_KEY, setAsJavaSet(Set(metadata)))
+
+    globalMetaData = new GlobalMetaData(globalMetaData.getSchema,
+      mergedMetadata, globalMetaData.getCreatedBy)
 
     val readContext = getReadSupport(configuration).init(
       new InitContext(configuration,