Align grouped first_value/last_value FILTER handling with shared Some(true) semantics

[kosiew]

Which issue does this PR close?

• Part of Centralize Aggregate FILTER Some(true) Semantics for Consistency and Correctness Hardening #22665

Rationale for this change

Grouped first_value and last_value used a local BooleanArray::value(idx) check when evaluating aggregate FILTER predicates. This reads only the value bit and does not account for NULL validity, which can incorrectly treat a NULL filter value with a true value bit as passing.

DataFusion's aggregate FILTER semantics require that only Some(true) passes; both Some(false) and None should reject the row. This change reuses the existing shared filter validity logic to ensure grouped first_value and last_value behave consistently with other aggregate implementations.

What changes are included in this PR?

• Made filter_to_validity publicly accessible from datafusion-functions-aggregate-common so it can be reused by grouped aggregate implementations.

• Updated FirstLastGroupsAccumulator to:

  • Precompute filter validity using filter_to_validity.
  • Evaluate FILTER predicates using the shared validity bitmap rather than BooleanArray::value(idx).

• Added targeted unit test helpers for grouped first_value / last_value accumulators.

• Added regression tests covering nullable FILTER predicates where the value bit is true but the validity bit is NULL:

  • test_first_group_acc_rejects_null_filter_with_true_value_bit
  • test_last_group_acc_rejects_null_filter_with_true_value_bit
  • test_first_group_acc_merge_rejects_null_filter_with_true_value_bit

Are these changes tested?

Yes.

This PR adds the following unit tests:

• test_first_group_acc_rejects_null_filter_with_true_value_bit
• test_last_group_acc_rejects_null_filter_with_true_value_bit
• test_first_group_acc_merge_rejects_null_filter_with_true_value_bit

These tests verify that rows with a NULL FILTER value do not pass even when the underlying boolean value bit is true, and that the same semantics are preserved in the merge path.

Are there any user-facing changes?

No user-facing changes are intended.

This is a correctness fix that aligns grouped first_value and last_value aggregate FILTER evaluation with existing DataFusion aggregate semantics for nullable filter predicates.

LLM-generated code disclosure

This PR includes LLM-generated code and comments. All LLM-generated content has been manually reviewed and tested.

[Nagato-Yuzuru]

The same bug in #22666. Looks like this PR covers that one too. Might be worth a sqllogictest on top since the repro was SQL. Here's mine if you want it (without the fix, g=1 returns 10 instead of NULL):

# Grouped first_value/last_value must apply aggregate FILTER with Some(true)
# semantics: a row passes only when the predicate is TRUE. Rows where the
# predicate evaluates to NULL or FALSE must be excluded.
#
# Rows per group (predicate is b < 1):
#   g=1: (a=10, b=NULL -> NULL), (a=20, b=2 -> FALSE)        => no rows pass
#   g=2: (a=30, b=0 -> TRUE), (a=40, b=NULL -> NULL),
#        (a=50, b=-5 -> TRUE)                                => a=30 and a=50 pass
#   g=3: (a=60, b=NULL -> NULL)                              => no rows pass
statement ok
CREATE TABLE first_last_filter_null_tests(g INT, a INT, b INT) AS VALUES
(1, 10, CAST(NULL AS INT)),
(1, 20, 2),
(2, 30, 0),
(2, 40, CAST(NULL AS INT)),
(2, 50, -5),
(3, 60, CAST(NULL AS INT));

# Groups 1 and 3 have no rows passing the filter -> NULL.
# Group 2 has a=30 and a=50 passing -> first_value ORDER BY a = 30.
query II
SELECT g, first_value(a ORDER BY a) FILTER (WHERE b < 1) AS fv
FROM first_last_filter_null_tests
GROUP BY g
ORDER BY g;
----
1 NULL
2 30
3 NULL

# Same groups via last_value: group 2 picks the largest passing a = 50.
query II
SELECT g, last_value(a ORDER BY a) FILTER (WHERE b < 1) AS lv
FROM first_last_filter_null_tests
GROUP BY g
ORDER BY g;
----
1 NULL
2 50
3 NULL

statement ok
DROP TABLE first_last_filter_null_tests;