add predict command

[FynnBe]

This PR

• updates the CLI (rewrite that now include a predict command)
• includes some other changes to allow loading statistical measures from json with pydantic validation

[thodkatz]

This is very cool :) I have already tested with affable-shark and the test input data provided by the model.

I have one general question, maybe I am missing something. Since we can have multiple input tensors and output tensors, that means we can have multiple .npy files. Why we don't store the tensors in one .npy file as an object, or maybe use the .npz, so we can can make our lifes easier dealing with one input and one output file?

[FynnBe]

I used .npy as a file format default here, but in general we support many file formats. Users might want use .h5, .tif .png... which the imageio library takes care of for us.

As there is no .npz equivalent across file formats I would like to keep this simple approach.
We might want to use .h5 as a default and we should add .zarr and possibly make that the default instead.
Using numpy is the most hacker friendly, but not the most user friendly.