Spread each segment of the segmented bitmask reduction across several blocks

[vuule]

Description

segmented_offset_bitmask_binop assigned one warp to each segment, so a segment's entire null mask was walked by 32 threads regardless of how long it is. For the struct null mask reduction this kernel exists to serve, a segment is one column's mask and its length is the table's row count, so reducing 192 masks of 500k rows ran on 6144 threads and reached roughly a tenth of achievable bandwidth. Blocks now cooperate on a segment when there are few segments, with each segment's null count accumulated across its blocks.

The existing parquet_read_fixed_width_struct is 5.2% to 6.2% faster end to end on an A100 80GB, a flat ~0.62 ms in all four configurations: the Parquet reader calls enforce_null_consistency on every struct hierarchy it produces, and that is this kernel. The existing parquet_read_decode -a data_type=STRUCT shows the same ~0.3 ms absolute saving, worth 1.4% to 2.4% there since string decode dominates that profile.

In isolation, segmented_bitmask_and on 8 to 512 segments of 100k to 1M bits is 23% to 98% faster. The existing bitmask axes stop at 128-bit masks, which fit in a handful of words and cannot show this, so a case with mask sizes in the range of real table row counts is added; those existing narrow configurations are unchanged within noise.

Independent of #23484, which removes the mask allocations on the same path. The two compose: 23484 calls inplace_segmented_bitmask_binop directly instead of segmented_bitmask_and, which is the same entry point reparallelized here and dispatches to the same kernel.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.