sql: preserve filters during distinct preprocessing

[Alignyx]

Fixes #158770.

Root cause

When every aggregate in a stage is DISTINCT, physical planning adds an early distinct processor keyed by the union of the aggregate argument columns. Filter columns were not included in that key. Rows with the same aggregate argument but different filter results could therefore be collapsed before each aggregate evaluated its own filter, causing multiple filtered DISTINCT aggregates over the same argument to observe the wrong subset of rows. Placeholders made the issue visible by preventing optimizer deduplication, but they were not the source of the error; two different literal filters failed in the same way.

How I tracked it down

I compared the reported prepared query with direct literal filters, reordered aggregates, a single filtered DISTINCT aggregate, non-DISTINCT controls, identical filters, different aggregate arguments, and COUNT plus SUM over the same DISTINCT argument. Row/vectorized and local/forced-distributed execution agreed on the failure, which placed it before either aggregation implementation. The physical plan contained both aggregate specs and both Boolean filter projections, but the shared all-DISTINCT preprocessing stage deduplicated only on the argument column. For a repeated argument, that stage could retain a false-filter row and discard the true-filter row before either aggregate ran.

Fix

The early distinct key now includes every non-nil aggregate filter column in addition to the existing union of argument columns. This preserves rows that are distinguishable by any later filter while retaining the shared preprocessing optimization and leaving individual aggregate implementations unchanged.

Test coverage

• Added a direct SQL regression with two literal filters over the same DISTINCT argument, expecting counts 2,1.
• Added a prepared regression with one literal and one parameterized filter, covering both equal thresholds (2,2) and different thresholds (2,1).
• Verified the focused regression in five local/fake-distributed row, vectorized, and disk-spilling configurations; ran the complete aggregate logic suite locally and fake-distributed, the full 16-shard SQL test target, and the five-node distributed aggregation suite containing the earlier distinct-preprocessing regression.

[blathers-crl]

Thank you for contributing to CockroachDB. Please ensure you have followed the guidelines for creating a PR.

My owl senses detect your PR is good for review. Please keep an eye out for any test failures in CI.

_{🦉 Hoot! I am a Blathers, a bot for CockroachDB. My owner is dev-inf.}