[ila]

Add handwritten stochastic TPC-DS queries and independent sampled-world references, with a runner for timing and checking the expanded workload. This separates benchmark development from the DuckDB 2.0 upgrade in #35.

Changes

• Preserve 66 handwritten AS queries and 66 per-world SQL references, the SF30 configuration, and the benchmark runner/validation additions.
• Document the distinction between compiled release behavior and per-world aggregate references.
• Keep the SUM accuracy unit correction directly in its SQL query.

This draft targets upgrade/duckdb-2.0 for a benchmark-only diff. It is intended to merge after #35 and be retargeted to main once that upgrade lands.

Validation and remaining work

The C++ runner builds against the cleaned upgrade; Python and JSON parse checks pass. A nonempty Q52 smoke succeeds in all four timing modes with five measured runs, and compiler/handwritten outputs match exactly.

Per-world validation remains WIP: setup unconditionally defines a macro using as_stddev_samp, which has no registered implementation. This prevents the AS per-world comparison even for queries that do not use standard deviation; Q17 also requires an actual STDDEV rewrite. The conditional-aggregate Q52 oracle matches. Resolve these paths before treating the per-world validation as complete.