[BUG] Incorrect output from partial-only averages with nulls

[kuhushukla]

Describe the bug
In partial only average aggregates having a columnar batch with one null entry for the sum column (averages are sent down as (sum,count) columns), results in wrong output for averages.

Steps/Code to reproduce bug
Here is a scala test to reproduce the issue and cn be run by adding it to HashAggregatesSuite.scala:

 IGNORE_ORDER_ALLOW_NON_GPU_testSparkResultsAreEqual(
    "PartMerge:avg_partOnly_null_corner_case", nullIntDf,
    execsAllowedNonGpu = Seq("HashAggregateExec", "AggregateExpression", "AttributeReference",
      "Alias", "Average", "Cast"),
    conf = partialOnlyConf,
    repart = 2) {
    frame => val result = frame.agg(avg("more_ints"))
      checkExecNode(result)
      result.explain()
      result
  }

def nullIntDf(session: SparkSession): DataFrame = {
    import session.sqlContext.implicits._
    Seq[(java.lang.Integer, java.lang.Integer)](
      (null, 15),
      (null, null)
    ).toDF("ints", "more_ints")
  }

Expected behavior
cpu avg = 15, gpu avg = null