Attach image transforms to model

Right now, some of the image transforms are hard coded. The transform configuration should be attached to the model, so model-dependent things like normalization come through with the model itself.

---
_Copied from [drivendataorg/zamba-wildlabs #67](https://github.com/drivendataorg/zamba-wildlabs/issues/67)_

Start by locating the hard-coded image transforms and the model construction or configuration entry points. Trace how normalization and other model-dependent settings are currently selected. Done means the transform configuration travels with the model rather than being maintained separately, with existing image-processing behavior preserved.

Written by the indexing model from the issue text.