ORDER BY ... LIMIT returns wrong rows for List/Struct sort keys with NULL elements #25998

Describe the bug

ORDER BY <list or struct column> LIMIT n can return the wrong rows when the sort key values contain NULL elements. The same query without LIMIT sorts correctly.

To Reproduce

SET datafusion.execution.target_partitions = 1;

create table t (l int[]) as values (make_array(1, NULL));
insert into t values (make_array(1, 2));

select l from t order by l limit 1;
-- [1, NULL]   (wrong)

select l from t order by l;
-- [1, 2]
-- [1, NULL]   (correct)

create table s as
select named_struct('a', 1, 'b', NULL::int) x
union all
select named_struct('a', 1, 'b', 2);

select x from s order by x limit 1;
-- {a: 1, b: NULL}   (wrong)

ORDER BY l DESC LIMIT 1 is also wrong (with the rows inserted in the opposite order), as are multi-column orderings such as ORDER BY a, l LIMIT 1 with ties on a, or ORDER BY l, b LIMIT 1.

Expected behavior

select l from t order by l limit 1 returns [1, 2], and select x from s order by x limit 1 returns {a: 1, b: 2}, i.e. the first row of the full ORDER BY result.

Additional context

The TopK operator builds its dynamic filter as col < threshold (or > for DESC) plus col = threshold terms. For nested types these comparisons go through compare_op_for_nested, which always uses SortOptions::default(), so NULL elements inside a list/struct are ordered differently from the sort itself (which follows the query's NULLS FIRST/LAST). With the threshold [1, NULL], [1, 2] < [1, NULL] evaluates to false and the correct row is filtered out.

TopK applies this filter to its own input as well as pushing it down, so setting datafusion.optimizer.enable_topk_dynamic_filter_pushdown = false (or enable_dynamic_filter_pushdown = false) does not avoid the bug.

#25957 is the same root cause in PiecewiseMergeJoinExec.