Correlated EXISTS on a CASE or UNION ALL column fails with "Can't compare arrays of different types" #26285

Describe the bug

A correlated EXISTS subquery fails at execution time with

Arrow error: Invalid argument error: Can't compare arrays of different types

when the outer column it references only gets its final type during type coercion, for example CASE WHEN ... THEN int_col ELSE 3 END (an INT branch first and a BIGINT one later) or a UNION ALL of an INT and a BIGINT column.

The subquery's filter never gets a cast: after type_coercion it is still u.c0 = outer_ref(s.k) although s.k has become Int64, so after decorrelation the join compares an Int64 key with an Int32 key.

To Reproduce

CREATE TABLE t(id INT, c1 INT) AS VALUES (1, 0), (100000, 0), (3, 5);
CREATE TABLE u(c0 INT) AS VALUES (1), (3), (100000);

SELECT s.k
FROM (SELECT CASE WHEN c1 = 0 THEN id ELSE 3 END AS k FROM t) s
WHERE EXISTS (SELECT 1 FROM u WHERE u.c0 = s.k);

Arrow error: Invalid argument error: Can't compare arrays of different types

The same happens with UNION ALL:

CREATE TABLE a(k INT) AS VALUES (1), (100000);
CREATE TABLE b(k BIGINT) AS VALUES (3);
-- u as above

SELECT s.k
FROM (SELECT k FROM a UNION ALL SELECT k FROM b) s
WHERE EXISTS (SELECT 1 FROM u WHERE u.c0 = s.k);

NOT EXISTS and a correlated scalar subquery such as (SELECT count(*) FROM u WHERE u.c0 = s.k) fail the same way. Putting the BIGINT branch first (CASE WHEN c1 <> 0 THEN 3 ELSE id END), or using a join instead of EXISTS, works. With small, dense keys such as 1, 2, 3 the default plan happens to return the right rows, but it still fails with SET datafusion.optimizer.prefer_hash_join = false.

Tested on main at 0566b55.

Expected behavior

Both queries return 1, 3 and 100000, as in PostgreSQL.

Additional context

For the CASE query, the subquery sees the outer column with its type from before coercion: arrow_typeof(s.k) is Int64 in the outer query, but inside the subquery arrow_typeof(s.k) = 'Int32' is true. (Expr::Case::get_type returns the type of the first non-null THEN branch.) The final plan joins k: Int64 with c0: Int32:

HashJoinExec: mode=CollectLeft, join_type=LeftSemi, on=[(k@0, c0@0)], schema=[k:Int64;N]
  ProjectionExec: expr=[CASE WHEN c1@1 = 0 THEN CAST(id@0 AS Int64) ELSE 3 END as k], schema=[k:Int64;N]
    DataSourceExec: partitions=1, partition_sizes=[1], schema=[id:Int32;N, c1:Int32;N]
  DataSourceExec: partitions=1, partition_sizes=[1], schema=[c0:Int32;N]