Fix Int4x2/UInt4x2 to MLFloat16 casting compilation errors on ARM64 Linux

[Copilot]

Problem

The build was failing on ARM64 Linux with compilation errors when trying to cast from Int4x2/UInt4x2 to Eigen::half:

/onnxruntime_src/build/Debug/Debug/vcpkg_installed/arm64-linux/include/Eigen/src/Core/MathFunctions.h:369:74: error: no matching function for call to 'onnxruntime::Int4x2Base<false>::Int4x2Base(const Eigen::half&)'
   369 |   EIGEN_DEVICE_FUNC static inline NewType run(const OldType& x) { return static_cast<NewType>(x); }
       |                                                                          ^~~~~~~~~~~~~~~~~~~~~~~

This error occurred specifically on ARM64 Linux builds but not on Windows x64.

Root Cause

The issue occurred because:

1. The generic TensorCaster template uses Eigen's casting: out_vector = in_vector.template cast<DstEigenCastType>();
2. EigenCastType<MLFloat16>::type maps to Eigen::half
3. Eigen doesn't know how to convert between Int4x2/UInt4x2 and Eigen::half
4. The existing specialized casting paths work on Windows x64 due to platform-specific optimizations, but fail on ARM64 Linux

Solution

Added explicit TensorCaster specializations that handle these conversions directly without going through Eigen:

• TensorCaster<Int4x2, MLFloat16> - Converts Int4x2 to MLFloat16 by unpacking 4-bit values and converting to float
• TensorCaster<UInt4x2, MLFloat16> - Converts UInt4x2 to MLFloat16 by unpacking 4-bit values and converting to float
• TensorCaster<MLFloat16, Int4x2> - Converts MLFloat16 to Int4x2 with proper clamping to signed int4 range [-8, 7]
• TensorCaster<MLFloat16, UInt4x2> - Converts MLFloat16 to UInt4x2 with proper clamping to unsigned int4 range [0, 15]

The specializations include:

• Proper value clamping using ToInt4ElementConverter helper functions
• Correct element packing/unpacking logic respecting Int4x2's nibble layout (elem 0 = low nibble, elem 1 = high nibble)
• Handling of odd-sized tensors with appropriate padding

Testing

• ✅ Manual compilation test confirms the fix resolves compilation issues
• ✅ Unit tests verify correct clamping logic and template specialization resolution
• ✅ No existing functionality is affected (0 lines deleted, 97 lines added)

This fix prevents the generic Eigen-based casting from being attempted for these specific type combinations while maintaining compatibility with existing optimized paths on all platforms.

---

💡 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.