LEAD/LAG with offset 0 and IGNORE NULLS returns the default for a NULL row, unlike the form without IGNORE NULLS #27284

Describe the bug

lead(v, 0, d) and lag(v, 0, d) return the current row's own value. Without IGNORE NULLS a NULL row stays NULL. With IGNORE NULLS the default d replaces it.

To Reproduce

SELECT id,
       lead(v, 0, -7) OVER (ORDER BY id) AS lead_plain,
       lead(v, 0, -7) IGNORE NULLS OVER (ORDER BY id) AS lead_ignore_nulls,
       lag(v, 0, -7) OVER (ORDER BY id) AS lag_plain,
       lag(v, 0, -7) IGNORE NULLS OVER (ORDER BY id) AS lag_ignore_nulls
FROM (VALUES (1, 10), (2, NULL), (3, 30)) AS t(id, v)
ORDER BY id;

On 55.2.0 (datafusion-python 55.0.0):

| id | lead_plain | lead_ignore_nulls | lag_plain | lag_ignore_nulls |
| 1  | 10         | 10                | 10        | 10               |
| 2  | NULL       | -7                | NULL      | -7               |
| 3  | 30         | 30                | 30        | 30               |

Expected behavior

Row 2 should be NULL in all four columns. Offset 0 refers to the current row itself, so there is no row to skip and nothing out of range that would call for the default. Apache Spark 4.0 returns the row's own value (NULL) for lead(v, 0, -7) IGNORE NULLS, the same as without IGNORE NULLS.

Additional context

In datafusion/functions-window/src/lead_lag.rs the two code paths disagree on this input as written:

• evaluate_all_with_ignore_null starts with if offset == 0 { return Ok(Arc::clone(array)) }, so a NULL stays NULL.
• WindowShiftEvaluator::evaluate ends with "Set the default value if ... ignore nulls mode and current value is null", which turns a NULL current row into the default when the offset is 0.

Offsets of 1 or more are not affected by this; the offset 0 case is only reached by an explicit 0.

Separately, #25471 (fixed on main by #25472) covers LEAD with an offset of 2 or more returning wrong rows across NULL gaps. The published 55.2.0 crate still has the old code for that.