IR: set-op and aggregate nullability, and the SQL top-k rewrite's output shape #657

Found by the DataFusion differential check of #539 (573 SQL queries run through DataFusion 54 and through the lowering). These are in the IR's schema derivation and canonicalization (already on main), not in the SQL frontend.

Problem

• Set operations take the left schema. SetOp derives its schema from left only (ir/non_asap.rs), so SELECT 1 AS x UNION ALL SELECT NULL is declared non-null but returns NULL. DataFusion declares it nullable. (feat(sql): lower SQL to unified operator and scalar IR #539 fixes the same issue for Concat, which its ROLLUP / CUBE / GROUPING SETS lowering uses.)
• Some aggregate results are declared non-null. stddev / variance / median / percentile results, and ClickHouse argMax / argMin, are NULL over empty input (e.g. stddev(a) … WHERE id = 1 with one row, or an empty group), but the aggregate schema derivation (pre_asap/agg_intent.rs, around the result-field construction) marks them non-null. The nullability fix in non_asap.rs covers only Sum / Avg / Min / Max.
• The top-k rewrite changes SQL output. canonicalize turns SELECT g, count(*) AS c … GROUP BY g ORDER BY c DESC LIMIT k into a one-column TopK aggregate instead of rows of (g, c), and extra projected columns (e.g. upper(g)) are dropped. This is the heavy-hitter rewrite and may be intended for planning, but the result shape no longer matches the SQL.

Possible fix

• SetOp: a column is nullable if it is nullable on either side.
• Aggregate schema derivation: mark these results nullable.
• Top-k: keep the SQL output shape (rows of the group key and the count) above the TopK, or apply the rewrite only where the output shape is preserved.