Implement Int4x2 and UInt4x2 casting support in ONNX Runtime CPU provider

[Copilot]

Summary

This PR implements comprehensive casting support for Int4x2 and UInt4x2 data types in the ONNX Runtime CPU provider, addressing the TODO comment in cast_op.cc and enabling 4-bit integer quantization workflows.

Changes Made

1. Updated Type Lists

• Changed cast operation type constraints from AllIRv9 to AllIRv10 to include Int4x2/UInt4x2 types
• Enables Int4x2/UInt4x2 as both source and destination types for cast operations

2. Core Casting Implementation

Added TensorCaster specializations for:

• Int4x2/UInt4x2 ↔ float: Primary conversions for quantization workflows
• Int4x2/UInt4x2 ↔ all numeric types: int8_t, uint8_t, int16_t, int32_t, int64_t, double, MLFloat16, BFloat16

3. Helper Functions and Templates

• ConvertFromInt4x2<Signed, DstType>(): Unpacks 4-bit values to destination types
• ConvertToInt4x2<Signed, SrcType>(): Converts and clamps values to 4-bit range with proper bounds checking
• Generic converter templates (Int4x2ToCaster, CasterToInt4x2) to eliminate code duplication

4. Key Features

• Proper nibble ordering: Follows "low nibble first, then high nibble" specification
• Value clamping: Signed Int4 values clamped to [-8, 7], unsigned UInt4 to [0, 15]
• Odd-sized tensor handling: Automatically pads incomplete pairs with zero
• Memory efficient: Leverages existing Int4x2/UInt4x2 infrastructure

Example Usage

// Int4x2 to float conversion
std::vector<Int4x2> int4_data = {Int4x2(3, -5), Int4x2(7, -8)};
// Unpacks to: [3.0, -5.0, 7.0, -8.0]

// Float to Int4x2 with clamping
std::vector<float> float_data = {3.7f, -5.2f, 10.0f, -15.0f, 2.1f};
// Clamps and packs to: [(3, -5), (7, -8), (2, 0)]  // Note: 10.0→7, -15.0→-8, padded with 0

Testing

Created comprehensive standalone tests validating:

• Correct unpacking of Int4x2/UInt4x2 to various numeric types
• Proper value clamping during conversion from larger types
• Correct handling of odd-sized tensors with zero padding
• Nibble ordering verification (val0→low nibble, val1→high nibble)

Compatibility

• No breaking changes: All existing cast operations continue to work
• Minimal footprint: Only 275 lines added, 3 lines modified
• Template-safe: Uses inheritance to avoid conflicting partial specializations
• Performance: Direct conversions without unnecessary intermediate steps

Resolves: TODO comment "Implement support for int4 and uint4" in cast_op.cc

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.