Make unet implementation scriptable

[FynnBe]

I started looking into improving affable-shark and its onnx weights in particular, which led me down the rabbit hole, so this PR makes the unet implementation mostly scriptable (with the small exception of the 'check_shape' validation).

In order to make it scriptable I

• added type annotations (in particular a few Final annotations to inform torch.jit about constants).
  Not all annotations are required, many I added for myself for a better overview and to help avoid unintended code changes.
• removed some of thereturn_outputs flags, changing the Encoder and Decoder interfaces slightly for static signatures needed for scripting.

I am not done testing the code with the affable-shark model example.

[FynnBe]

I will try to allow scale_factor to be List[float] again, this seems to be used outside of unet.py...