[FEA] Support Group-By on Array[String]

[viadea]

I wish we can support group-by on Array[String].

Minimum repro:

import org.apache.spark.sql.types.{ArrayType, StringType, StructType}
val arrayData = Seq(
    Row("John",List("apple","orange","banana")),
    Row("David",List("apple","orange","banana")),
    Row("Harry",List("apple","other"))
)

val arraySchema = new StructType().add("name",StringType).add("fruits", ArrayType(StringType))

val df = spark.createDataFrame(spark.sparkContext.parallelize(arrayData),arraySchema)
df.write.mode("overwrite").format("parquet").save("/tmp/testparquet")
val df2=spark.read.parquet("/tmp/testparquet")
df2.createOrReplaceTempView("df2")
spark.sql("select fruits from df2 group by fruits").show()

Not-supported-messages:

 !Exec <HashAggregateExec> cannot run on GPU because ArrayTypes or MapTypes in grouping expressions are not supported
    @Expression <AttributeReference> fruits#71 could run on GPU
    @Expression <Alias> cast(fruits#71 as string) AS fruits#76 could run on GPU
      @Expression <Cast> cast(fruits#71 as string) could run on GPU
        @Expression <AttributeReference> fruits#71 could run on GPU
    !Exec <ShuffleExchangeExec> cannot run on GPU because not all partitioning can be replaced; Columnar exchange without columnar children is inefficient
      !Partitioning <HashPartitioning> cannot run on GPU because hash_key expression AttributeReference fruits#71 (ArrayType(StringType,true) is not supported)
        @Expression <AttributeReference> fruits#71 could run on GPU
      !Exec <HashAggregateExec> cannot run on GPU because ArrayTypes or MapTypes in grouping expressions are not supported

[revans2]

I filed rapidsai/cudf#10181 as the CUDF dependency that we need for this.

[revans2]

I split up the original CUDF issue into two and added rapidsai/cudf#10184 for sorting of lists of strings. It is not required to do the initial work, but would really be nice to have.

[revans2]

In CUDF the ask is being listed as a special case of rapidsai/cudf#8039 so this is actually likely to be more generic in that the key could be lists of lists/etc.

[sameerz]

Depends on adding murmur3 hashing for list columns issue rapidsai/cudf#10378