[SPARK-20038][SQL] FileFormatWriter.ExecuteWriteTask.releaseResources() implementations to be re-entrant

## What changes were proposed in this pull request?

have the`FileFormatWriter.ExecuteWriteTask.releaseResources()` implementations  set `currentWriter=null` in a finally clause. This guarantees that if the first call to `currentWriter()` throws an exception, the second releaseResources() call made during the task cancel process will not trigger a second attempt to close the stream.

## How was this patch tested?

Tricky. I've been fixing the underlying cause when I saw the problem [HADOOP-14204](https://issues.apache.org/jira/browse/HADOOP-14204), but SPARK-10109 shows I'm not the first to have seen this. I can't replicate it locally any more, my code no longer being broken.

code review, however, should be straightforward

Author: Steve Loughran <[email protected]>

Closes #17364 from steveloughran/stevel/SPARK-20038-close.

Author: steveloughran

sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/FileFormatWriter.scala

@@ -324,8 +324,11 @@ object FileFormatWriter extends Logging {
 
     override def releaseResources(): Unit = {
       if (currentWriter != null) {
-        currentWriter.close()
-        currentWriter = null
+        try {
+          currentWriter.close()
+        } finally {
+          currentWriter = null
+        }
       }
     }
   }
@@ -459,8 +462,11 @@ object FileFormatWriter extends Logging {
 
     override def releaseResources(): Unit = {
       if (currentWriter != null) {
-        currentWriter.close()
-        currentWriter = null
+        try {
+          currentWriter.close()
+        } finally {
+          currentWriter = null
+        }
       }
     }
   }