Added ARRAY data type (fixed-sized lists)

[yayen-lin]

Add ARRAY (fixed-size list) data type support

Summary

• Maps DuckDB's ARRAY type to cuDF LIST in the type system.
• cuCascade handles the H2D transfer and memory management for the LIST column.
• list_offset_fixup.cpp/.hpp: a temporary workaround (see #147).

Changes

Type system

• src/include/helper/logical_type.hpp: add the ARRAY logical type.
• src/helper/type_conversions.cpp: ARRAY to/from DuckDB conversion.
• src/include/cudf/cudf_utils.hpp: map ARRAY to/from cuDF LIST.

Scan / read-back

• src/op/scan/duckdb_scan_task.cpp: lower a DuckDB ARRAY vector into a cuDF LIST column.
• src/op/result/host_table_chunk_reader.{hpp,cpp}: read a cuDF LIST column back into a DuckDB ARRAY vector.
• src/op/scan/duckdb_native_metadata.cpp: classify ARRAY as a nested type in the native-scan viability switch.

LIST offset fixup

• src/helper/list_offset_fixup.cpp, src/include/cudf/list_offset_fixup.hpp: recast LIST offsets back to INT32 after cuCascade's H2D promotes them to INT64.
• src/include/pipeline/batch_lock_utils.hpp: apply the fixup after the H2D conversion.
• CMakeLists.txt: add list_offset_fixup.cpp and tests to the build.

Tests

• test/cpp/helper/test_logical_type.cpp: ARRAY logical type tests.
• test/cpp/helper/test_cudf_utils.cpp: ARRAY to cuDF LIST mapping tests.
• test/cpp/integration/test_gpu_execution_array.cpp: end-to-end ARRAY scan tests (deferred, see below).

What's deferred

• End-to-end GPU scan of ARRAY columns is deferred because it's out of scope for this PR. The DuckDB-native scan path rejects nested types, so transparent GPU execution of an ARRAY query isn't viable yet.

[mbrobbel]

Thanks @yayen-lin

[yayen-lin]

Hi Matthijs, thank you for taking your time to review this PR, I really appreciate it. Please take a look at my response and the new commit if you have a chance.

[yayen-lin]

Hey Matt good afternoon,

I was looking at the failed ci job, I thought that was a native scan path where Sirius would read directly from DuckDB's on-disk storage.

Could you let me know if I should get that to work, or whether it's fine to leave it as is for now?

Thanks,
Andy

[bwyogatama]

Can we add PR description and also @kevkrist has solved that issue i believe. You should rebase and that will probably solve the CI

[yayen-lin]

Thank you Bobbi, I just rebased with latest code on dev.

Also just a heads up, I added [.] to the test (test_gpu_execution_array.cpp) since GPU decoding of the ARRAY type is out of the scope for this PR.

[bwyogatama]

Okay so the biggest problem with this PR is that you are using the CPU duckdb scan path, which will be deprecated. You need to use the duckdb scan path with GPU cause otherwise we will lose this feature as soon as this is deprecated. DuckDB GPU scan path is served through duckdb_native_gpu_ingestible

[yayen-lin]

Got it, thank you for your feedback Bobbi, I will work on that!