Use cuDF to count the total deleted row count for cuDF based Delta reader for deletion vectors

[sdrp713]

Fixes #14628

Related CUDF changes: NVIDIA/cudf#23978

Description

Update the cuDF based Delta deletion vector reader to use the new JNI/libcudf GPU row-count API.

The reader uses the GPU count when it needs the number of surviving rows for:

• Zero-column scans such as SELECT count(*).
• Populating partition columns before attaching them to a ColumnarBatch.

This removes host-side deletion-vector bitmap traversal from these paths while preserving the existing row-group metadata calculation.

Testing

• 48 targeted Delta deletion-vector integration tests passed (test_delta_deletion_vector_read); remaining selected cases were skipped by existing runtime guards.
• Validated PERFILE, COALESCING, and MULTITHREADED readers.

Performance

100 million rows, 5 million deleted rows, 2 warmups, and 10 measured runs. The cudf-spark and JNI changes were tested together.

Reader/query	Baseline mean	PR mean	Speedup
PERFILE / count_star	0.747868s	0.702228s	1.06×
PERFILE / partition_aggregate	1.017174s	0.983110s	1.03×
COALESCING / count_star	0.484052s	0.432153s	1.12×
COALESCING / partition_aggregate	0.713125s	0.666528s	1.07×
MULTITHREADED / count_star	0.407790s	0.368274s	1.11×
MULTITHREADED / partition_aggregate	0.592371s	0.585036s	1.01×

Checklists

Documentation

• Updated for new or modified user-facing features or behaviors
• No user-facing change

Testing

• Added or modified tests to cover new code paths
• Covered by existing tests
  (Please provide the names of the existing tests in the PR description.)
• Not required

Performance

• Tests ran and results are added in the PR description
• Issue filed with a link in the PR description
• Not required

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge; no outstanding correctness, resource-lifecycle, or repository-rule violation was established.

Summary

• Uses GPU deletion-vector cardinality computation under the RAPIDS retry and semaphore mechanisms.
• Preserves serialized bitmap ownership through spillable host buffers.
• Cancels outstanding bitmap loads and restores thread interruption when asynchronous preparation fails.
• Precomputes row counts used by zero-column scans and partition-column population.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Load serialized deletion vectors] --> B[Spillable host buffers]
  B --> C[Group adjacent files by partition value]
  C --> D[cuDF computeNumDeletedRows]
  D --> E[Compute surviving rows]
  E --> F[Decode filtered table]
  E --> G[Populate partition columns]

Loading

[gerashegalov]

A future may succeed before it was cancelled. On this thread future.get will throw a cancelled exception and we will never get a hold of the gpuBitmap to close

[jihoonson]

We should batch the call per partition instead of per file. We don't really need to track the deleted row count per file. We just need to know it per partition. The cudf c++ API already supports the batch computation, but we may need to modify the JNI interface.

[sdrp713]

Cudf PR for this here: NVIDIA/cudf#24047

[gerashegalov]

Fix builds for OSS Delta shims