feat(lower): nested SQL query functions via derived tables (#27)

[zzylol]

What & why

Adds nested query function support for SQL — the SQL counterpart of the PromQL nesting in #28 — closing the SQL part of #27.

SQL's lower_plan already recurses through the DataFusion LogicalPlan, so a nested aggregate/filter/projection chain lowered. The gap was derived tables / inline views (FROM (SELECT …) t): the SubqueryAlias arm only accepted a TableScan (or another alias) as input, so an aggregate-over-aggregate had nowhere to go.

SELECT MAX(s) FROM (SELECT service, SUM(bytes) AS s FROM metrics GROUP BY service) t
SELECT AVG(p) FROM (SELECT service, approx_percentile_cont(latency, 0.9) AS p FROM metrics GROUP BY service) t
SELECT t.service, t.s FROM (SELECT service, SUM(bytes) AS s FROM metrics GROUP BY service) t WHERE t.s > 100

How

• lower_plan's SubqueryAlias arm now lowers an arbitrary inner plan recursively (lower_plan already handles every node a sub-SELECT produces). The alias is dropped; a qualified outer ref (t.col) resolves by bare name against the derived output schema — the same Qualified → bare-name fallback the converter's column resolution already uses for joins. No L3 IR / converter / Binder changes needed.
• Subquery-valued expressions in a predicate (x > (SELECT …), IN (SELECT …), EXISTS (…)) need a subquery node in the L2 expression IR (+ a correlated/uncorrelated decision), so they're rejected cleanly with an explicit message instead of the generic catch-all.

Tests

• derived_table_aggregate_over_aggregate_nests — outer Max over inner Sum, schema derivation total across the boundary.
• derived_table_outer_avg_over_inner_percentile — exact Avg over approximate Quantile (per-node sketch-vs-exact is an L4 decision).
• filter_over_derived_aggregate_resolves_alias_column — WHERE t.s > 100 resolves the qualified alias column.
• scalar_subquery_in_predicate_is_rejected — pinned rejection.
• Full workspace suite green; cargo clippy --all-targets clean.

Scope / follow-ups

Scalar / IN / EXISTS subqueries in predicates (correlated subqueries) remain rejected — tracked in the updated #27. Relates to #28 (PromQL nesting); the two are independent and don't conflict.

Closes the SQL part of #27.

🤖 Generated with Claude Code