Add DLPack and numpy array protocols to OrtValue Python wrapper

[Rishi-Dave]

Summary

• Add __dlpack__(), __dlpack_device__(), and from_dlpack() to the Python OrtValue class, enabling zero-copy tensor sharing with PyTorch, JAX, NumPy, and other DLPack-compatible frameworks.
• Add __array__() protocol so np.array(ort_value) and np.asarray(ort_value) work directly.
• The C binding (C.OrtValue) already supported these methods — this PR exposes them through the public Python wrapper class.

Motivation

Fixes #24071

The Python OrtValue wrapper class did not expose the DLPack or numpy array protocols, even though the underlying C binding had full support (added in #23110). Users had to access the internal _ortvalue attribute to use __dlpack__ or from_dlpack, which is fragile and undocumented.

With this change, standard Python interop patterns work out of the box:

# DLPack protocol — zero-copy to/from PyTorch, JAX, etc.
torch_tensor = torch.from_dlpack(ort_value)
np_array = np.from_dlpack(ort_value)
ort_value = OrtValue.from_dlpack(torch_tensor)

# Numpy array protocol
np_array = np.array(ort_value)
np_array = np.asarray(ort_value)

Changes

• onnxruntime/python/onnxruntime_inference_collection.py: Added 4 methods to OrtValue:
  • __dlpack__(*, stream=None) — returns a DLPack capsule (delegates to C.OrtValue.__dlpack__)
  • __dlpack_device__() — returns (device_type, device_id) tuple
  • from_dlpack(data) — classmethod accepting any __dlpack__-compatible object or raw capsule, with automatic bool tensor detection via is_dlpack_uint8_tensor
  • __array__(dtype=None, copy=None) — numpy array protocol, compatible with NumPy 1.x and 2.x
• onnxruntime/test/python/onnxruntime_test_python.py: Added 9 test cases covering:
  • __dlpack__ and __dlpack_device__ on the Python wrapper
  • from_dlpack with OrtValue→OrtValue, numpy→OrtValue, and raw capsule→OrtValue
  • np.from_dlpack(ort_value) interop
  • np.array(), np.asarray(), and dtype conversion via __array__
  • Full round-trip across multiple dtypes (float32, float64, int32, int64)

Test Plan

• All 9 new tests pass locally against pip-installed onnxruntime with DLPack enabled
• Existing test_ort_value_dlpack and test_ort_value_dlpack_zero_size tests unaffected
• ruff check and ruff format --check pass on both modified files
• CI validation