[POC] Back libcudf scalars with Arrow-compatible column storage

[GregoryKimball]

Background

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

• Scalar validity is stored separately from column null-mask metadata. Consequently, scalar::is_valid() synchronizes with the device (#8064), and scalar validity cannot safely be interpreted as a column bitmask (#22757).
• Fixed-width, decimal, string, list, and struct scalars use type-specific storage instead of column-compatible data, masks, offsets, and children.
• Algorithms duplicate scalar/column paths or materialize temporary one-row columns (#18858, #10892).
• Device code uses a separate hierarchy of typed scalar views that lacks a generic, const-correct interface (#13160, #4365).
  These are different logical abstractions, but they do not need different physical storage. A scalar can own the same Arrow-compatible layout as a one-row column while retaining scalar-specific constructors and accessors.
  This implements the central storage proposal from #19465.

Summary

• Backs numeric, chrono, fixed-point, string, list, and struct scalars with one-row cudf::column storage.
• Adds allocation-free const and mutable scalar column views.
• Makes host null-count metadata authoritative, so scalar::is_valid() no longer synchronizes.
• Provides mutable scalar views for value writes while requiring validity to be established explicitly.
• Preserves existing typed scalar constructors and value accessors as compatibility facades.
• Migrates compiled binary operations, JIT literals, search, sequence, copying, reductions, shared iterators, string kernels, and decoded-key dictionary operations to standard column device views.
• Removes scalar conversion allocations, typed device-view adapters, and const_casts from migrated consumers.
• Retains owning one-row materialization for internally synthesized JIT temporaries whose lifetime would otherwise be too short.

Benchmark results

For order-32 FLOAT32 polynomials, the JIT path improves most because it avoids per-literal one-row materialization. AST and binary-op paths show smaller gains because they already used lightweight scalar adapters.

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.