Hadoop ingestion ignores useDefaultValueForNull=false for metrics with rollup enabled

When using index_hadoop to ingest a parquet file with null values in a metric column, the useDefaultValueForNull=false is not respected when rollup is enabled.
longSum aggregations for dimensions where the metric values are all null, are calculated as 0 instead of null.

Ingesting the same file with index_parallel from local storage or natively from S3, results in the correct behaviour and the longSum aggregations are calculated as null.

### Affected Version

0.19.0

### Description

The parquet file has 3 columns - name (string), age (int), index (int) and was created with the following Spark code:

```
val dfdata = Seq(Row("Dan",35, null), Row("Dan",34, null), Row("John",20, 2), Row("Mike",30, 9), Row("Sam",40, 0), Row("Tom", 17, 1))

val dfSchema = List(StructField("name", StringType, true), StructField("age", IntegerType, true), StructField("index", IntegerType, true))

val df = spark.createDataFrame(spark.sparkContext.parallelize(dfdata), StructType(dfSchema))

```

As you can see, the "index" column is null for all records with name = 'Dan', thus I would expect the longSum aggregation of "index" to be null for these rows.

The parquet file itself is attached.

The ingestion spec used for ingestion via hadoop:

```
{
"type": "index_hadoop",
"spec": {
"ioConfig": {
"type": "hadoop",
"inputSpec": {
"type": "static",
"paths":"s3a:///part-00000-5933233a-9db6-4bbb-8529-fa9687c9b2f1-c000.gz.parquet",
"inputFormat": "org.apache.druid.data.input.parquet.DruidParquetInputFormat"
}
},
"tuningConfig" : {
"type": "hadoop"
},
"dataSchema": {
"dataSource": "parquet_test_hadoop",
"parser": {
"type": "parquet",
"parseSpec": {
"format": "parquet",
"dimensionsSpec": {
"dimensions": ["name"],
"dimensionExclusions": []
},
"timestampSpec": {
"missingValue": "2010-01-01T00:00:00.000Z",
"format": "auto",
"column": "timestamp"
}
}
},
"metricsSpec":[
{"name":"sum_idx", "type":"longSum","fieldName":"index"}
],
"granularitySpec": {
"type": "uniform",
"queryGranularity": "DAY",
"rollup": true,
"segmentGranularity": "DAY"
}
}
}
}
```

The ingestion spec used to ingest natively from S3:

```
{
"type": "index_parallel",
"spec": {
"dataSchema": {
"dataSource": "parquet_test_s3",
"timestampSpec": {
"column": "timestamp",
"format": "auto",
"missingValue": "2010-01-01T00:00:00.000Z"
},
"dimensionsSpec": {
"dimensions": ["name"]
},
"metricsSpec": [
{
"type": "longSum",
"name": "sum_idx",
"fieldName": "index"
}
],
"granularitySpec": {
"type": "uniform",
"segmentGranularity": "DAY",
"queryGranularity": "DAY",
"rollup": true
}
},
"ioConfig": {
"type": "index_parallel",
"inputSource": {
"type": "s3",
"uris": [
"s3:///part-00000-5933233a-9db6-4bbb-8529-fa9687c9b2f1-c000.gz.parquet"
],
"properties": {
"accessKeyId": {
"type": "default",
"password": ""
},
"secretAccessKey": {
"type": "default",
"password": ""
}
}
},
"inputFormat": {
"type": "parquet"
}
},
"tuningConfig": {
"type": "index_parallel"
}
},
"dataSource": "parquet_test_s3"
}
```

useDefaultValueForNull is not respected for hadoop ingestion:


useDefaultValueForNull is respected for native ingestion:

[part-00000-5933233a-9db6-4bbb-8529-fa9687c9b2f1-c000.gz.parquet.gz](https://github.com/apache/druid/files/5709576/part-00000-5933233a-9db6-4bbb-8529-fa9687c9b2f1-c000.gz.parquet.gz)

Start by reproducing the issue with the attached Parquet file and the provided index_hadoop ingestion spec, then compare it with the index_parallel S3 path. Trace the Hadoop Parquet input path and rollup handling for null metric values. Done means useDefaultValueForNull=false produces null longSum results for dimensions whose metric values are all null, matching native ingestion.

Written by the indexing model from the issue text.