[FEA][STORY] Support ANSI Operators using JIT

[lamarrr]

Description

This issue tracks the implementation of ANSI SQL semantics in the cuDF JIT expression engine.

The goal of this work is to enable ANSI-compliant operator execution through CUDF JIT code generation, allowing expression evaluation to remain fully device-side while preserving Spark-compatible overflow, nullability, precision, and error semantics.

This effort builds on the initial prototype introduced in #22224 and continues the broader work described in #21676.

The implementation focuses on:

• Extending transforms with ANSI-aware operators
• Preserving ANSI overflow and precision semantics during code generation
• Supporting nullable execution paths and TRY_* semantics
• Enabling decimal-aware arithmetic and rescaling
• Maintaining parity with existing libcudf expression semantics where applicable

Implemented Operators

Nullability / Conditional Semantics

• NULLIFY_IF
• COALESCE
• IF_ELSE

ANSI Arithmetic

• ANSI_ADD
• ANSI_SUB
• ANSI_MUL
• ANSI_DIV
• ANSI_MOD
• ANSI_ABS
• ANSI_NEG

ANSI TRY Semantics

• ANSI_TRY_ADD
• ANSI_TRY_SUB
• ANSI_TRY_MUL
• ANSI_TRY_DIV
• ANSI_TRY_MOD
• ANSI_TRY_ABS
• ANSI_TRY_NEG

Precision / Decimal Semantics

• ANSI_PRECISION_CHECK
• ANSI_TRY_PRECISION_CHECK
• RESCALE
• CAST_TO_DEC32
• CAST_TO_DEC64
• CAST_TO_DEC128

Integer / Floating-Point Casts

• CAST_TO_B8
• CAST_TO_I8
• CAST_TO_I16
• CAST_TO_U8
• CAST_TO_U16
• CAST_TO_U32
• CAST_TO_U64
• CAST_TO_F32

Bitwise Operators

• BIT_SHIFT_LEFT
• BIT_SHIFT_RIGHT

Implementation Notes

The ANSI operators are implemented as first-class JIT expression operators and lower directly into generated CUDA/PTX code paths. Special care is taken to:

• Preserve overflow detection semantics across signed arithmetic operations
• Propagate validity masks consistently for nullable execution
• Support decimal scale adjustments and precision enforcement in-device
• Avoid unnecessary host-side branching or fallback execution
• Maintain compatibility with Spark ANSI mode behavior

TRY_* operators follow ANSI semantics by converting runtime arithmetic failures into null outputs rather than raising execution errors.

Work Breakdown

The original prototype in #22224 was intentionally split into smaller reviewable units:

• [FEA] ANSI SQL Operator JIT Support (1) : Refactor ROW IR #22511
• [FEA] ANSI SQL Operator JIT Support (2) : Implement Operator Library #22514
• [] [FEA] ANSI SQL Operator JIT Support (3) : Implement IR for ANSI and other JIT Extensions #22602

Future Work

Potential follow-on work includes:

• Additional ANSI comparison and predicate operators
• Expanded decimal intrinsic support
• LTO-IR/PTX codegen for operator fusion
• Consolidation of duplicated AST/JIT operator semantics