SQL: shapes DataFusion supports that the lowering rejects #658

Found by the DataFusion differential check of #539. Each of these is rejected with an explicit error, so nothing is computed wrongly, but DataFusion supports them.

Rejected shapes

• ClickHouse scalar builtins such as startsWith, match, positionCaseInsensitive, splitByChar fail with "unregistered scalar function": the IR function registry (ir/scalar.rs) does not know them, although the frontend registers their DataFusion stubs and a code comment says registering the stub is the entire fix. toDate / toStartOf* likely have the same gap (not run: no timestamp data in the check).
• INTERSECT / EXCEPT [ALL] and LEFT SEMI / LEFT ANTI JOIN are rejected (frontend-sql/src/sql/mod.rs, set-op and join arms), although the IR has these set-op kinds and the Semi / Anti join kinds.
• Two window functions sharing one window spec are rejected; the error message suggests this is deliberate, but the shape is common.
• Others: PARTITION BY <expression>; IS [NOT] DISTINCT FROM (also in JOIN ON); IS TRUE / IS NOT TRUE / IS UNKNOWN; ||; SIMILAR TO; ~; substr; ntile; percent_rank; bool_and; array_agg; aggregate first_value; SUM(DISTINCT); GROUPING(); two-argument uniqExact; ClickHouse quantile(q)(x).

Rejections the #539 description already calls deliberate are not listed (DISTINCT ON, GROUPS frames, non-literal LIMIT, correlated subqueries other than a filter-conjunct EXISTS, count(nullable) under ROLLUP).

Follow-up

Support the shapes the workloads need, starting with the ClickHouse builtins, which the BGP corpora use.