Do not import torch outside of klongpy/backends/torch_backend.py.

All torch-specific code must be confined to the torch backend module. Other modules (interpreter, compiler, monads, dyads, adverbs) must not import torch directly. When torch functionality is needed elsewhere:

• Use klong._backend.np.ndarray to check for tensors (resolves to torch.Tensor on torch backend, numpy.ndarray on numpy)
• Use hasattr(klong._backend, '_torch_backend') to detect the torch backend
• Use sys.modules.get('torch') to get the torch module reference when the torch backend is already active
• Use backend methods (e.g., klong._backend.np.cumsum) for operations that differ between backends

This keeps torch as an optional dependency and prevents import-time failures when torch is not installed.

Run the full test suite with:

pip install -e ".[full]"
python -m unittest discover tests