Describe the bug
After upgrading to PyTorch 2.8.0, compilation involving LabelTensor fails due to incorrect inference of runtime_type in the torch compilation pipeline. This results in a RuntimeError during the backward pass.

To Reproduce
Upgrade to PyTorch 2.8.0 and run:

pytest tests/test_solver/test_pinn.py

You might also want to upgrade torchvision.

Expected behavior
All tests should pass successfully, and compilation should handle LabelTensor without metadata mismatches.

Output

RuntimeError: 
E               During the backward, we encountered a tensor subclass where we guessed its
E               metadata incorrectly.
E               
E               Expected metadata: None, expected type: <class 'torch.Tensor'>
E               
E               Runtime metadata: None, runtime type: <class 'pina.label_tensor.LabelTensor'>
E               
E               shape: torch.Size([1, 1])
E               To fix this, your tensor subclass must implement the dunder method __force_to_same_metadata__.

Additional context

• Implementing __force_to_same_metadata__ in LabelTensor did not resolve the issue.
• This might be related to changes introduced in the compilation pipeline in PyTorch 2.8.
• The issue was not observed in PyTorch 2.7.x.