Bugfix desc null fix + add tpcds fixture

[ran-yuan-rui]

Summary

Split out from #806 per @wmalpica review comment.

This PR contains two independent pieces:

1. Fix GPU ORDER BY / TOP-N correctness for DESC ... NULLS FIRST/LAST.
2. Add a reusable SF0.01 TPC-DS DuckDB fixture for integration / plan-translation tests.

RCA: DESC NULL ordering bug

DuckDB SQL treats NULLS FIRST / NULLS LAST as the final, absolute placement of NULLs in the result:

• DESC NULLS FIRST => NULLs first, then values descending
• DESC NULLS LAST => values descending, then NULLs last

Sirius previously mapped SQL null ordering directly to cuDF:

NULLS_FIRST -> cudf::null_order::BEFORE
NULLS_LAST  -> cudf::null_order::AFTER

That is correct for ASC, but wrong for DESC: cuDF applies null_order in its sort-direction frame, so DESC effectively reverses the NULL placement. As a result, GPU ORDER BY k DESC NULLS LAST could place NULLs first, disagreeing with DuckDB CPU.

The same naive mapping was duplicated across the distributed sort path:

• ORDER_BY
• SORT_SAMPLE
• SORT_PARTITION
• MERGE_SORT
• TOP_N

There was one additional TOP-N issue: cudf::top_k_order has no null_order parameter, so single-key TOP-N on a nullable key could select the wrong top-k rows before any final sorting step.

Fix

• Added src/include/op/cudf_sort_order.hpp as the single source of truth:
  • to_cudf_order(...)
  • to_cudf_null_order(...)
• to_cudf_null_order(...) flips BEFORE / AFTER for descending keys so SQL NULLS FIRST/LAST matches DuckDB semantics.
• Routed all sort/top-n operators through the helper.
• For single-key TOP-N:
  • keep cudf::top_k_order fast path when the key has no NULLs
  • use sort_by_key + slice when the key is nullable, so NULL placement is honored

TPC-DS fixture

Update (reworked per @mbrobbel's review): Dropped the committed 12 MB tpcds.duckdb binary. The fixture is now generated at build time via an opt-in CMake target tpcds-fixture (runs generate_tpcds_duckdb.sh into ${CMAKE_BINARY_DIR}/test-fixtures/, SF 0.01 via DuckDB's dsdgen). Run cmake --build . --target tpcds-fixture once before the TPC-DS integration tests.

Tests

Added test_gpu_execution_order_nulls.cpp:

• ORDER BY
• TOP-N single-key
• TOP-N multi-key
• ASC/DESC × NULLS FIRST/LAST

The test materializes GPU output once, adds observed_position to make ordering differences visible, and compares against DuckDB CPU with bidirectional EXCEPT ALL.

Validation:

pre-commit run --all-files
pixi run -e default make release
build/release/extension/sirius/test/cpp/sirius_unittest "[nulls]"