[kvlonge]

Why

A SQL macro meant for several grains had to be written once per grain, and COLUMN OF split one
input across arguments that had to agree with it.

Changes

• RELATION (...) declares slots: typed single columns and untyped COLUMNS groups.
• Calls map slots with rel.slot := column or rel.slot := (a, b). Groups must be mapped.
• Reads follow the slot kind, never the mapping:
  • alias.@rel.slot reads through the alias; a group's prefix applies to every column.
  • A plain group read always stands as bare column names, so later CTEs need no alias. A name
    two joined relations share fails with B003, as hand-written SQL does.
  • A plain single-slot read always reads its column through the passed relation.
  • Reads in USING (...) are always bare. Forwarded slots keep their inner kind.
• Capture check: a plain group read in a subquery (including LATERAL) fails with P017 when a
  relation there other than the passed one has one of its columns, or its columns are unknown.
• Group rules and call checks (columns, types, AS outside expression subqueries, P018
  clashes with defined or bare columns) have exact messages.
• Rules lint a slot read without taking the slot's name as its column.
• COLUMN OF is removed; use relation slots.
• The duckshop benchmarks use slots. Docs and the bundled skill are updated.

BREAKING CHANGE: COLUMN OF SQL macro arguments are removed; use relation slots.

Verification

• Rust tests for expansion by kind, group rules and call errors.
• DuckDB E2E: a macro reading @weekly.keys plainly in PARTITION BY, later CTEs and GROUP BY
  compiles at two grains to SQL byte-identical to two per-grain macros, with the same rows.
  Capture, LATERAL, forwarding, USING, subquery AS and clash cases, ambiguous joins,
  lineage, Rules, typing, --since.
• Both duckshop benchmarks compile with no diagnostics.
• make test, make check-ci, fensu check and the hygiene scan pass.