[BUG] collect_set on a LIST<LIST<INT32>> with all null entries returns the wrong result type

[jlowe]

Describe the bug
Performing a collect_set aggregation on a lists column of lists of ints returns the wrong resulting scalar type when the input lists column contains all nulls. Instead of returning a LIST<LIST<INT32>> it returns a LIST<INT32>. When there are non-null entries in the input it properly returns LIST<LIST<INT32>>.

Steps/Code to reproduce bug
Compile and run the following program.

#include <cudf_test/debug_utilities.hpp>
#include <cudf_test/column_wrapper.hpp>
#include <cudf/aggregation.hpp>
#include <cudf/reduction.hpp>
#include <cudf/scalar/scalar.hpp>
#include <iostream>

using LCW = cudf::test::lists_column_wrapper<int>;

void debug(cudf::lists_column_view const& lists) {
  std::cout << "Lists size: " << lists.size() << " null count: " << lists.null_count() << std::endl;
  auto child = lists.child();
  int id = static_cast<int>(child.type().id());
  std::cout << "Lists child type: " << id << " size: " << child.size() << " null count: " << child.null_count() << std::endl;
  if (child.type().id() == cudf::type_id::LIST) {
    child = cudf::lists_column_view(child).child();
    id = static_cast<int>(child.type().id());
    std::cout << "Lists child child type: " << id << " size: " << child.size() << " null count: " << child.null_count() << std::endl;
  }
}

void collect_set(cudf::lists_column_view const& lists) {
  std::cout << "Before reduction:" << std::endl;
  debug(lists);
  auto agg = cudf::make_collect_set_aggregation<cudf::reduce_aggregation>(cudf::null_policy::EXCLUDE, cudf::null_equality::UNEQUAL, cudf::nan_equality::ALL_EQUAL);
  auto r = cudf::reduce(lists.parent(), *agg, cudf::data_type{cudf::type_id::LIST});
  std::cout << "After reduction:" << std::endl;
  auto rl = dynamic_cast<cudf::list_scalar*>(r.get());
  debug(rl->view());
}

int main(int argc, char** argv) {
  auto valids = cudf::detail::make_counting_transform_iterator(0, [](auto i) { return i != 0; });
  cudf::lists_column_view lists = cudf::test::lists_column_wrapper<int32_t, int32_t>{ {LCW{{1, 2, 3}}, LCW{{1, 2, 3}}}, valids };
  std::cout << "=== With null list and non-null list ===" << std::endl;
  collect_set(lists);
  std::cout << "=== With only null list ===" << std::endl;
  lists = cudf::test::lists_column_wrapper<int32_t, int32_t>{ {LCW{{1, 2, 3}}}, valids };
  collect_set(lists);
  return 0;
}

This produces the following output:

Before reduction:
Lists size: 2 null count: 1
Lists child type: 24 size: 1 null count: 0
Lists child child type: 3 size: 3 null count: 0
After reduction:
Lists size: 0 null count: 0
Lists child type: 24 size: 0 null count: 0
Lists child child type: 3 size: 0 null count: 0
=== With only null list ===
Before reduction:
Lists size: 1 null count: 1
Lists child type: 24 size: 0 null count: 0
Lists child child type: 3 size: 0 null count: 0
After reduction:
Lists size: 0 null count: 0
Lists child type: 3 size: 0 null count: 0

Note how the result types change based on whether the input has all nulls or not:

After reduction:
Lists size: 0 null count: 0
Lists child type: 24 size: 0 null count: 0
Lists child child type: 3 size: 0 null count: 0

vs.

After reduction:
Lists size: 0 null count: 0
Lists child type: 3 size: 0 null count: 0

Expected behavior
The collect_set aggregation should consistently return the same output type for a given input type, regardless of the input data for that type. In this specific case, it should have returned LIST<LIST<INT32>> instead of LIST<INT32>.