[PyOV] Fix from_torchvision with a torch.Tensor input_example on NumPy 2

[Somay-kousis]

Details:

• What a user sees: PreprocessConverter.from_torchvision(model, transform, input_example=torch.rand(1, 3, H, W)) raises ValueError: Unable to avoid copy while creating an array as requested. on NumPy 2. The docstring lists torch.Tensor as a supported input_example. np.ndarray and PIL.Image examples work, and nothing changes for them.
• Cause: _get_shape_layout_from_data reads the shape with np.array(input_example, copy=False). NumPy 2 changed copy=False from "copy only if needed" to "never copy", and torch.Tensor.__array__ takes no copy argument, so NumPy raises. The bindings allow numpy<2.6.0.
• Fix: np.asarray(input_example), which is the NumPy 1 meaning of copy=False (NumPy's error message suggests the same replacement). Only .shape is used.
• New test test_torch_tensor_input_example in tests/layer_tests/py_frontend_tests/test_torchvision_preprocessor.py: a torch.Tensor example with Normalize, compared against the torch model. The existing tests only pass PIL.Image examples.

Tickets:

• Fixes [Bug]: [PyOV] from_torchvision fails with a torch.Tensor input_example on NumPy 2 #38937

AI Assistance:

• AI assistance used: yes
• AI tools helped find the bug and write the fix and the test; I reviewed every line. Validation on macOS arm64 with the 2026.5.0 nightly wheel, numpy 2.5.4, torch 2.11, torchvision 0.26, using this branch's torchvision_preprocessing.py:
  • test_torch_tensor_input_example: passes. With master's file it fails with the ValueError above.
  • test_torchvision_preprocessor.py: 15 passed, 14 skipped (the existing ARM skips for Ticket 114816).
  • The repro from [Bug]: [PyOV] from_torchvision fails with a torch.Tensor input_example on NumPy 2 #38937 with np.ndarray, PIL.Image and torch.Tensor examples: all three build the model.