Add support for UINT2/INT2 data types

[Copilot]

Add support for UINT2/INT2 data types

Following the pattern from PR onnx#128 (FLOAT8E8M0 support), this PR adds comprehensive support for INT2/UINT2 data types that were added to ONNX in PR #7446.

Changes Made:

• Add INT2 and UINT2 enum values to _enums.py (values 25 and 26)
• Add numpy type mappings for ml_dtypes.int2 and ml_dtypes.uint2
• Add type properties (bitwidth=2, signed/unsigned, integer type)
• Add short names ("i2" and "u2")
• Update _core.py to handle INT2/UINT2 in packed tensor lists and numpy view functions
• Add pack_2bitx4 and unpack_2bitx4 functions in _type_casting.py for 2-bit packing
• Update serde.py to include INT2/UINT2 in serialization sets and handle 2-bit unpacking
• Update tensor_adapters.py for torch tensor support (mappings added, but tests excluded as PyTorch doesn't support creating tensors with these types yet)
• Add comprehensive tests in _core_test.py, _enums_test.py, serde_test.py
• Update PackedTensor to support 2-bit types alongside 4-bit types
• Update tensors.md documentation to include INT2/UINT2
• All tests passing

Testing:

All existing tests pass, and new tests have been added to verify:

• Enum values and type properties
• Numpy type conversions
• Tensor creation and serialization
• Pack/unpack operations for 2-bit data
• Torch dtype mappings (though PyTorch doesn't yet support creating tensors with these types)
• Serde round-trip operations

Notes:

• While PyTorch has torch.int2 and torch.uint2 dtype definitions, it doesn't support creating tensors with these types yet, so torch tensor tests are excluded
• The tensor_adapters.py file includes the mappings for future compatibility when PyTorch adds full support

Original prompt

Helpe me create PR for this issue: onnx#250

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.

[vraspar]

@copilot Do i need to add anything to tensors.md to support int2/uint2

[Copilot]

@copilot Do i need to add anything to tensors.md to support int2/uint2

Yes, I've updated the documentation in tensors.md to include INT2/UINT2. The changes are in commit aabcb44:

• Updated the section title to include "int2"
• Added [U]INT2 to the list of non-native NumPy dtypes
• Updated the unpacking note to include uint2/int2
• Updated initialization constraints to include int2 and uint2

This follows the same documentation pattern used for INT4/UINT4.