Check if inputs are valid according to input shape parametrization

[constantinpape]

I don't think that we check whether a model input is valid according to the shape parametrization when running prediction.
I think that we should check this, but only throw a warning if it is not valid, since the model might still be able to process the image.
E.g. because we have the constrained that the min_shape must be big enough to crop away the halo, so smaller shapes might actually pass without issue.
This came up here: stardist/stardist#171 (comment)

[constantinpape]

Thus is done now.