NPE in ParquetInputFormat.getSplits when no .parquet files exist

```Java
JavaSparkContext context = ...
JavaRDD rdd1 = context.parallelize(ImmutableList. of());
SQLContext sqlContext = new SQLContext(context);
StructType schema = DataTypes.createStructType(ImmutableList.of(DataTypes.createStructField("col1", DataTypes.StringType, true)));
DataFrame df = sqlContext.createDataFrame(rdd1, schema);
String url = "file:///tmp/emptyRDD";
df.saveAsParquetFile(url);

Configuration configuration = SparkHadoopUtil.get().newConfiguration(context.getConf());
JobConf jobConf = new JobConf(configuration);
ParquetInputFormat.setReadSupportClass(jobConf, RowReadSupport.class);
FileInputFormat.setInputPaths(jobConf, url);
JavaRDD rdd2 = context.newAPIHadoopRDD(
jobConf, ParquetInputFormat.class, Void.class, Row.class).values();
rdd2.count();

df = sqlContext.createDataFrame(rdd2, schema);
url = "file:///tmp/emptyRDD2";
df.saveAsParquetFile(url);

FileInputFormat.setInputPaths(jobConf, url);
JavaRDD rdd3 = context.newAPIHadoopRDD(
jobConf, ParquetInputFormat.class, Void.class, Row.class).values();
rdd3.count();
```

The NPE happens here:
```Java
java.lang.NullPointerException
at parquet.hadoop.ParquetInputFormat.getSplits(ParquetInputFormat.java:263)
at parquet.hadoop.ParquetInputFormat.getSplits(ParquetInputFormat.java:245)
at org.apache.spark.rdd.NewHadoopRDD.getPartitions(NewHadoopRDD.scala:95)
```

This stems from ParquetFileWriter.getGlobalMetaData returning null when there are no footers to read.

**Reporter**: [Paul Nepywoda](https://issues.apache.org/jira/secure/ViewProfile.jspa?name=pnepywoda)

_{**Note**: *This issue was originally created as [PARQUET-294](https://issues.apache.org/jira/browse/PARQUET-294). Please see the [migration documentation](https://issues.apache.org/jira/browse/PARQUET-2502) for further details.*}

Beginne mit ParquetInputFormat.getSplits und verfolge, wie ParquetFileWriter.getGlobalMetaData behandelt wird, wenn keine Footer gefunden werden. Reproduziere die bereitgestellte empty-RDD-Sequenz und überprüfe, dass sowohl rdd2.count() als auch rdd3.count() ohne eine NullPointerException abgeschlossen werden.

Vom Indexierungsmodell aus dem Issue-Text verfasst.