[ila]

Why

DuckDB 2.0 keeps some optimized one-sided outer joins as RIGHT joins. The regular N-term compiler accepted INNER and LEFT joins but rejected RIGHT, so Spark's gold.fact_market_history fell back to a 15-term inclusion-exclusion program during SF100 batch 2.

Change

Accept RIGHT joins in the same one-sided outer-join path as LEFT joins. The existing term builder and NULL-transition handling already orient themselves by the null-supplying side; no new delta rule is needed.

Evidence

• An exact TPC-DI-shaped fact_market_history compile shrinks from 112,470 to 43,301 SQL bytes, with 14 to 9 UNION ALL occurrences while preserving the RIGHT join.
• Extended the existing regular N-term SQL test with both match-transition directions and full EXCEPT ALL equality after a mixed INSERT/DELETE/UPDATE batch.
• test/sql/left_join_regular_nterm.test: 60 assertions pass.
• test/integration/test_regular_nterm_compiled.py: passes.
• Spark/Delta integration with the new native binary: RIGHT-join MV is bag-equal before and after one mixed refresh.

This PR is stacked on #147 because the benchmark currently pins its head. It does not change DuckLake row-ID upsert logic.