Fix scalar as column view for null scalars

[mhaseeb123]

Description

Closes #22757. Supersedes #22759.

This PR fixes scalar_as_column_view to handle null input input scalar by allocating a single-element bitmask for them (kept alive using the aux column). Valid scalars are still directly just converted to column views with nullptr as their nullmask.

The newly added tests surface the memcheck error when compute-sanitizer is run with --rmm_mode=cuda.

Credits: @simoneves for the original PR.

Checklist

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

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR fixes a memory safety bug in cuDF's scalar-to-column-view conversion where null-aware binary operations incorrectly reinterpret scalar device bool validity data as a column bitmask, causing out-of-bounds reads. The fix explicitly materializes proper null masks for invalid scalars and omits masks for valid ones in both fixed-width and string_view specializations, validated by new null-aware fixed-point test cases.

Changes

Scalar null-mask handling in binary operations

Layer / File(s)	Summary
Include and fixed-width scalar null-mask conversion cpp/src/binaryop/compiled/binary_ops.cu	Added cudf/detail/null_mask.hpp header. Reworked scalar_as_column_view for fixed-width types to branch on validity: invalid scalars now create a proper single-element null mask via create_null_mask and return an auxiliary column retaining that mask; valid scalars construct column_view without a null mask. Removed prior invalid reinterpret_cast of scalar validity storage.
String_view scalar null-mask conversion cpp/src/binaryop/compiled/binary_ops.cu	Updated cudf::string_view scalar specialization with the same validity branching: valid scalars produce column_view without null mask while keeping offsets child alive; invalid scalars create all-null mask, attach it to the column_view, and return auxiliary column retaining both mask and offsets child.
Test coverage for null-aware scalar operations cpp/tests/binaryop/binop-compiled-fixed_point-test.cpp	Added FixedPointBinaryOpNullMaxColumnScalar test validating NULL_MAX with non-null scalar produces nullable output with all rows valid. Added FixedPointBinaryOpNullEqualsNullColumnScalar test validating NULL_EQUALS with invalid scalar produces boolean results that are true only where column is also null.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~25 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title 'Fix scalar as column view for null scalars' directly summarizes the main change: addressing how null scalars are handled in the scalar_as_column_view function.
Linked Issues check	✅ Passed	The PR addresses all coding requirements from issue #22757: materializing a proper 1-element null bitmask for null scalars, keeping it alive via auxiliary column, and maintaining nullptr nullmask for valid scalars. Tests verify null-aware binary ops (NULL_MAX, NULL_EQUALS) work correctly.
Out of Scope Changes check	✅ Passed	All changes are directly related to fixing scalar_as_column_view for null scalar handling as specified in issue #22757; no unrelated modifications detected.
Description check	✅ Passed	The PR description clearly relates to the changeset, explaining the fix to scalar_as_column_view for null scalars and referencing the added tests.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[bdice]

Looks good to me.

The newly added tests surface the memcheck error when compute-sanitizer is run with --rmm_mode=cuda.

Apologies, I should have suggested this earlier. I hope this didn't take too long to discover. OOB accesses are often reproducible only with the cuda_memory_resource. (Which is why we set it in our sanitizer tests.)

[bdice]

Just a reminder that we need to be careful with the data model here. A non-nullable column and a nullable column with valid data are not identical states. They can round-trip differently to/from various file formats.

If I recall correctly, cudf scalars don't distinguish "valid" from "non-nullable" states today, and thankfully this doesn't have much impact on I/O because we only read/write columns and not scalars, but it's important to keep that in mind.

No action needed, just want to build awareness.

[mhaseeb123]

A non-nullable column and a nullable column with valid data are not identical states. They can (and do) round-trip differently to/from various file formats

Learned this the hard way back when I was a new hire while working on round trips with Arrow 😄

[mhaseeb123]

If I recall correctly, cudf scalars don't distinguish "valid" from "non-nullable" states today, and thankfully this doesn't have much impact on I/O because we only read/write columns and not scalars, but it's important to keep that in mind.

Yup, this makes sense. Good for a temporary column view state but not when writing/reading.

[mhaseeb123]

I hope this didn't take too long to discover.

I had just asked Claude to write the smallest possible repro (the test in the PR) and run it under compute-sanitizer. This was the first thing it did when the test didn't fail. All this happened while I was working on something else 😄

[simoneves]

LGTM
Tested in Velox
All decimal tests run clean under compute-sanitizer --tool memcheck

[mhaseeb123]

/merge