take kernel on List array introduces nulls instead of empty lists

[jonmmease]

Describe the bug
In looking into apache/datafusion#4828, I came across what I think is an issue with the take kernel when operating on List arrays. It seems that when a taken element of the list array is an empty list, the value becomes null in the result of take.

I think this is a bug, but let me know if it's the intended behavior.

To Reproduce
Here is an example that stats with the list array: [[0], [], [1]] and takes elements [0, 1, 2].

use arrow::array::{Int32Array, Int32Builder, ListBuilder};
use arrow::compute::take;

#[test]
fn test_take() {
    let mut val_builder = Int32Builder::new();
    let mut list_builder = ListBuilder::new(val_builder);
    list_builder.values().append_value(0);
    list_builder.append(true);
    list_builder.append(true);
    list_builder.values().append_value(1);
    list_builder.append(true);
    let list = list_builder.finish();

    println!("list: {:?}", list);
    let indices = Int32Array::from(vec![0, 1, 2]);
    let taken = take(&list, &indices, None).unwrap();

    println!("taken: {:?}", taken);
}

output

list: ListArray
[
  PrimitiveArray<Int32>
[
  0,
],
  PrimitiveArray<Int32>
[
],
  PrimitiveArray<Int32>
[
  1,
],
]

taken: ListArray
[
  PrimitiveArray<Int32>
[
  0,
],
  null,
  PrimitiveArray<Int32>
[
  1,
],
]

Expected behavior
I would expect the result to be the same as the input ([[0], [], [1]]), but instead the result is [[0], null, [1].

Additional context
It looks like this behavior is the result of

arrow-rs/arrow-select/src/take.rs

Lines 706 to 710 in 2d2d0a3

	if window[0] == window[1] {
	// offsets are equal, slot is null
	bit_util::unset_bit(null_slice, i);
	null_count += 1;
	}

[tustvold]

Agreed this behaviour is wrong, I'd be happy to review a fix

[tustvold]

label_issue.py automatically added labels {'arrow'} from #3473