ci: sweep TPC-H/TPC-DS verification over partition counts and empty tables

[andygrove]

Which issue does this PR close?

N/A — test-infrastructure follow-up to the correctness bugs fixed in #2185, #2186, #2188, and #2194.

Rationale for this change

The recent correctness bugs shared a pattern: a distributed plan rewrite that was structurally plausible but semantically wrong, triggered by runtime conditions our CI never produces. The --verify oracle (cluster result vs single-process DataFusion) runs only at --partitions 16 on full TPC-H SF10 / TPC-DS SF1 data, which never yields an empty intermediate stage and never exercises low partition counts.

This PR makes CI exercise exactly those conditions.

What changes are included in this PR?

• A new --empty-tables <t1,t2,...> flag on both benchmark drivers. It registers the named tables as zero-row Parquet mirrors (schema inferred from the real data) in both the Ballista session and the --verify oracle, so verification remains a valid comparison while every query touching an emptied table produces empty intermediate stages through the distributed planner. On the TPC-H driver it requires --format parquet and is incompatible with --expected.
• TPC-H workflow: matrix extended from 3 to 6 legs — the existing SF10 legs at 16 partitions, plus SF1 AQE-on legs at 1 and 2 partitions, plus an SF1 AQE-on empty-tables leg (the emptied set rotates by day of week across lineitem / orders / supplier,nation). Job/step timeouts added so a hang fails fast, and cluster logs now upload on timeout as well as failure.
• TPC-DS workflow: matrix extended from 2 to 4 legs, adding target_partitions=1 and target_partitions=2 legs alongside the existing 16-partition task-packing legs.
• benchmarks/README.md documentation for the new flag.

Notes for maintainers:

• The empty-tables leg exists to surface latent empty-stage planner bugs. If it exposes one, please --skip the affected query with an issue link rather than reverting the leg.
• The rotation is keyed on day-of-week, so re-running a red empty-tables leg after UTC midnight tests a different table set and may go green without a fix. The leg logs Empty tables for this run: ... — reproduce locally with that printed set via --empty-tables.

Are there any user-facing changes?

No. The new CLI flag is on the benchmark binaries only; no library or cluster behavior changes.