Allow backbone to be any functional, preprocessor any callable

[mattdangerw]

The main change this allows is passing arbitrary functional models as a backbone to a task and arbitrary callables as preprocessing for a task. This should allow a lot more flexible API usage.

• Passing Rescaling layer or a Pipeline as preprocessing for a ImageClassifier task.
• Running augmentation on the GPU by retracing the backbone with augmentation layers included.
• Doing more custom surgery of a backbone.

I also moved our saving routines to a PresetSaver class, but this is really just to be parallel with the PresetLoader we added for transformers/timm conversion. You can still override save_to_preset on a model if you need to. Probably more cleanups to do.

[mattdangerw]

Merging this! The last failure is unrelated, and has to do with image preprocessing on the torch backend and gpu <> cpu errors. I will fix up on the other PR.