[TPC-DS] Fix paired sales/returns table generation

[kevinjqliu]

This affects only TPC-DS because TPC-H has no sales/returns pair concept.

Follow-up to #354 and #391; part of #218.

Requirement

For each sales/returns pair:

• Sales only → generate and output only sales.
• Returns only → generate and output only returns.
• Both → generate both in one shared pass and output both.

Generated work and output files must exactly match the selected tables.

Issues

1. Selecting one table produces both

Affected formats: DAT and CSV.

Selecting only sales or only returns generates and writes both tables.

2. Selecting both traverses the shared source twice

Affected format: Parquet.

Selecting both tables writes the correct outputs but generates each independently instead of sharing one source pass.

Behavior matrix

Using store_sales and store_returns as the representative pair:

Format	Sales only	Returns only	Both
DAT¹	❌ Outputs both	❌ Outputs both	✅ One pass, both outputs
CSV	❌ Outputs both	❌ Outputs both	✅ One pass, both outputs
Parquet	✅ Sales only	✅ Returns only	❌ Two source passes

¹ DAT is the TPC-DS equivalent of TBL.

Reproduction

Issue 1: DAT and CSV output both tables

# DAT: each command incorrectly produces both paired files
cargo run -p tpcgen-cli --bin tpcgen-cli tpcds dat --scale-factor=1 --tables=store_sales -o "$(mktemp -d)"
cargo run -p tpcgen-cli --bin tpcgen-cli tpcds dat --scale-factor=1 --tables=store_returns -o "$(mktemp -d)"

# CSV: each command incorrectly produces both paired files
cargo run -p tpcgen-cli --bin tpcgen-cli tpcds csv --scale-factor=1 --tables=store_sales -o "$(mktemp -d)"
cargo run -p tpcgen-cli --bin tpcgen-cli tpcds csv --scale-factor=1 --tables=store_returns -o "$(mktemp -d)"

Selecting both already works:

cargo run -p tpcgen-cli --bin tpcgen-cli tpcds dat --scale-factor=1 --tables=store_sales,store_returns -o "$(mktemp -d)"
cargo run -p tpcgen-cli --bin tpcgen-cli tpcds csv --scale-factor=1 --tables=store_sales,store_returns -o "$(mktemp -d)"

Issue 2: Parquet traverses the shared source twice

Single-table selection works:

cargo run -p tpcgen-cli --bin tpcgen-cli tpcds parquet --scale-factor=1 --tables=store_sales -o "$(mktemp -d)"
cargo run -p tpcgen-cli --bin tpcgen-cli tpcds parquet --scale-factor=1 --tables=store_returns -o "$(mktemp -d)"

Selecting both produces the correct files but traverses the shared source twice:

cargo run -p tpcgen-cli --bin tpcgen-cli tpcds parquet --scale-factor=1 --tables=store_sales,store_returns -o "$(mktemp -d)"

Acceptance criteria

• DAT/CSV single-table selection generates and writes only the selected table.
• DAT/CSV paired selection continues to generate both in one pass.
• Parquet paired selection generates both in one shared pass.
• Cover store, catalog, and web pairs across every affected format.

Repeated selection is already handled by #391 (TPC-DS) and #392 (TPC-H) and is outside this issue.