Fix null handling for structs min and arg_min in groupby, groupby scan, reduction, and inclusive_scan

[ttnghia]

When finding min, max, arg_min and arg_max for structs in groupby, groupby scan, reduction and inclusive_scan operations, null struct rows should be excluded from the operation (but the null rows of its children column are not). The current implementation for structs wrongly includes nulls at all levels, producing wrong results for min and arg_min operations.

This PR fixes that. In particular, null rows at the children levels are still being handled by the old way (nulls are smaller than non-null elements), but handling nulls at the top parent column level is modified such that:

• nulls are considered as larger than all other non-null rows, if finding for min and arg_min, or
• nulls are considered as smaller than all other non-null rows, if finding for max and arg_max.

[jrhemstad]

However, when finding min struct elements, null are considered last after all non-null elements.

Says who?

[ttnghia]

However, when finding min struct elements, null are considered last after all non-null elements.

Says who?

Says Spark:

scala> df.show
+-----+
|value|
+-----+
| null|
|    1|
| null|
|    2|
+-----+


scala> df.sort("value").show
+-----+
|value|
+-----+
| null|
| null|
|    1|
|    2|
+-----+


scala> df.selectExpr("min(value)").show
+----------+
|min(value)|
+----------+
|         1|
+----------+

And struct:

scala> df.show
+---+------------+                                                              
|  b|           a|
+---+------------+
|  1|        null|
|  1|{null, null}|
|  1|      {1, 2}|
|  1|   {null, 1}|
|  1|   {1, null}|
|  1|      {1, 1}|
|  2|        null|
+---+------------+

scala> df.selectExpr("min(a)").show()
+------------+
|      min(a)|
+------------+
|{null, null}|
+------------+

[jrhemstad]

Says Spark:

libcudf is not Spark.

Here's what the docs say happen with nulls in reduce and scan:

cudf/cpp/include/cudf/reduction.hpp

Line 41 in 2e95fb1

* The null values are skipped for the operation.

cudf/cpp/include/cudf/reduction.hpp

Lines 75 to 76 in 2e95fb1

	* The null values are skipped for the operation, and if an input element
	* at `i` is null, then the output element at `i` will also be null.

[ttnghia]

Here's what the docs say happen with nulls in reduce and scan:

Got it. I'll update the PR. Thanks.
The same results will be achieved, but we will look at the situation differently.

Update: Done.

[nvdbaranec]

Is num_rows here the same as table.num_rows(). table_device_view::num_rows() is callable on the host, so you could drop the parameter.

[ttnghia]

Done.

[nvdbaranec]

It would be a little cleaner to pass the aggregation to the comparison_binop_generator and have it do the is_min_op logic internally. That way the user doesn't have to type it out every time.

[ttnghia]

Done.

[davidwendt]

The null_at(2) does not seem to match the NULL comments above. Perhaps I'm reading this wrong.

[ttnghia]

That child column has null at idx 2. The following nulls are at the parent column which are specified below (nulls_at({3, 4, ...})). Sorry for the confusion.
Edit: I have added more comments to clarify it.

[ttnghia]

@gpucibot merge