Bring model.to in line with PyTorch module.to (bc)

[RunDevelopment]

This makes the behavior of our model.to the same as PyTorch's module.to. This is a breaking change, because we previously interpreted string arguments differently.

I had to add typing_extensions as a dependency, because I needed Self (py >=3.11) to get the return type of to right.

Other changes:

• I noticed that the return types of cuda and cpu were wrong and changed them to Self.
• I noticed that the inferred return types of eval and train did not show up in the docs, and add Self as the return type.