Add weights format 'python'

[FynnBe]

in today's later part of the meeting the idea of formulating specialized transformations as bioimage.io models came up.
Our old ideas about a dedicated transformation spec could be revived by just slightly adapting the model spec, but even more promising and much easier would be to add a "python" weight format (or "pickle") to run a special python model/transformation (e.g. stardist postprocessing).

All fields of the current model spec are valid, only for the "weights" field the wording is slightly off, but some pure python models/transformations might make use of pickled weights. Other than that the model would behave just like a pytorch_state_dict model with source and kwargs.

cc @oeway @constantinpape

[oeway]

Related to this, one could then use libraries such as scikit-image (e.g. watershed) or albumentations or cellpose for providing generic "python models" that can be used as a custom pre-/post-processing. The only change we need is to support a special type of pre and post processing named model, then in the kwrags we provide an url/doi to the python model.