torch.Tensor support for methods in numpy submodule

[agoscinski]

In #62 we introduce tox testing environments for metatensor-core with numpy arrays and with torch tenors. For now in tests/equisolve_tests/numpy/utilities.py we set use_torch to False, so all tests are not run using numpy as array backend. But we probably also want to support torch tensors for modules like Ridge and SparseKRR. Internally everything will be converted to numpy arrays (which we eventually change also at some point).

This would go hand in hand with a renaming of the numpy submodule to something less specific. One could use the term shallow for shallow ML methods like linear and kernel models. Or we split it up into kernel_ridge and linear_model as scikit-learn does it.