Inspect input and output dtypes together with value ranges for all stages

At least we should maybe mention this in `logger.debug`.

Anyway, by default the pipeline should do as few castings as possible but at the same time we need to make sure
the ranges and dtypes are compatible with a given stage. Most notably we should make sure the input for the interpolation stage is correct since running neural network inference on images with values between `0 - 255` is probably a bad ide