[SPARK-25669][SQL] Check CSV header only when it exists

## What changes were proposed in this pull request?

Currently the first row of dataset of CSV strings is compared to field names of user specified or inferred schema independently of presence of CSV header. It causes false-positive error messages. For example, parsing `"1,2"` outputs the error:

```java
java.lang.IllegalArgumentException: CSV header does not conform to the schema.
 Header: 1, 2
 Schema: _c0, _c1
Expected: _c0 but found: 1
```

In the PR, I propose:
- Checking CSV header only when it exists
- Filter header from the input dataset only if it exists

## How was this patch tested?

Added a test to `CSVSuite` which reproduces the issue.

Closes #22656 from MaxGekk/inferred-header-check.

Authored-by: Maxim Gekk <maxim.gekk@databricks.com>
Signed-off-by: hyukjinkwon <gurwls223@apache.org>

sql/core/src/main/scala/org/apache/spark/sql/DataFrameReader.scala

@@ -505,7 +505,8 @@ class DataFrameReader private[sql](sparkSession: SparkSession) extends Logging {
     val actualSchema =
       StructType(schema.filterNot(_.name == parsedOptions.columnNameOfCorruptRecord))
 
-    val linesWithoutHeader: RDD[String] = maybeFirstLine.map { firstLine =>
+    val linesWithoutHeader = if (parsedOptions.headerFlag && maybeFirstLine.isDefined) {
+      val firstLine = maybeFirstLine.get
       val parser = new CsvParser(parsedOptions.asParserSettings)
       val columnNames = parser.parseLine(firstLine)
       CSVDataSource.checkHeaderColumnNames(
@@ -515,7 +516,9 @@ class DataFrameReader private[sql](sparkSession: SparkSession) extends Logging {
         parsedOptions.enforceSchema,
         sparkSession.sessionState.conf.caseSensitiveAnalysis)
       filteredLines.rdd.mapPartitions(CSVUtils.filterHeaderLine(_, firstLine, parsedOptions))
-    }.getOrElse(filteredLines.rdd)
+    } else {
+      filteredLines.rdd
+    }
 
     val parsed = linesWithoutHeader.mapPartitions { iter =>
       val rawParser = new UnivocityParser(actualSchema, parsedOptions)

sql/core/src/test/scala/org/apache/spark/sql/execution/datasources/csv/CSVSuite.scala

@@ -1820,4 +1820,10 @@ class CSVSuite extends QueryTest with SharedSQLContext with SQLTestUtils with Te
     checkAnswer(spark.read.option("multiLine", true).schema(schema).csv(input), Row(null))
     assert(spark.read.csv(input).collect().toSet == Set(Row()))
   }
+
+  test("field names of inferred schema shouldn't compare to the first row") {
+    val input = Seq("1,2").toDS()
+    val df = spark.read.option("enforceSchema", false).csv(input)
+    checkAnswer(df, Row("1", "2"))
+  }
 }