[POC] Back libcudf scalars with one-row column storage

[GregoryKimball]

Background

cudf::scalar and cudf::column currently use incompatible device layouts:

• Scalar validity is a device bool, so host-side is_valid() requires synchronization ([FEA] cudf::scalar cache is_valid to reduce amount of stream syncs #8064) and cannot safely serve as a column bitmask ([BUG] Possible bug with fixed-point scalar handling in some binary operators #22757, Fix scalar as column view for null scalars #22773).
• Fixed-width, decimal, string, list, and struct scalars each own type-specific storage instead of Arrow-compatible data, masks, offsets, and children.
• Algorithms must duplicate scalar/column paths or materialize temporary one-row columns ([FEA] Add overloads for row comparator that automatically handle comparisons between rows of a table with a scalar #10892, Refactor scalar/column conversion logic in binary ops #18858).
• Scalars require a separate hierarchy of typed device views, with no generic or const-correct interface ([FEA] Add Generic scalar device view #13160, [BUG] get_scalar_device_view doesn't support const scalar  #4365).
• Constructor, validity, stream, and nested-scalar behavior is inconsistent ([FEA] Scalar factory functions do not let you provide a validity bool. #8633, [FEA] Stream-preserving constructors for scalar class and its derived classes #14333).
• scalar_column_view already demonstrates the desired representation. Reusing one-row columns in AST/JIT avoids thousands of copies and has shown speedups up to 27x ([BUGFIX] Fix handling of scalar inputs in transform and row_ir #21443, [FEA] Add support for scalar column views in AST expressions and JIT execution #23615), but ordinary scalars still use the old storage model.

Sumary

• Unifies numeric, chrono, fixed-point, string, list, and struct scalar storage with column layout.
• Adds allocation-free const and mutable one-row column views.
• Makes scalar validity host-authoritative and non-synchronizing, matching the column null-count contract.
• Requires explicit reconciliation when device code modifies a scalar null mask.
• Preserves existing typed scalar constructors and accessors as compatibility facades.
• Removes scalar conversion allocations, typed device-view adapters, and const casts
• Retains owning materialization for internally synthesized JIT temporaries to prevent dangling views.

This substantially addresses:

• #8064: scalar::is_valid() no longer synchronizes.
• #18858: compiled binary operations consume scalar storage directly.
• #4365: migrated consumers are const-correct and no longer cast away scalar constness.
• #10892 and #10893: scalars can participate directly in column-oriented algorithms without one-row materialization.
• #13160: standard column device views now replace typed scalar device views in representative consumers.

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]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.