[Feature] Make the cellfinder training data generation workflow clearer

I think it should be made clearer that the general workflow is:
- Run cellfinder
- Load curation widget
- Create *new* training data layers
- Add points to these new layers
- Extract cubes
- Train
- Run cellfinder
- Repeat curation/training as necessary, either by adding to the existing training dataset, or creating new ones

A user was curating the cellfinder results directly. This meant:
- Cube generation took forever
- Retraining the network didn't actually impact performance much

No documentation file or test is named. Start by locating the cellfinder training and curation workflow documentation in this website repository, then compare it with the sequence described in the issue. Done means the guidance clearly distinguishes new training-data layers from curated cellfinder results and explains when to repeat training.

Written by the indexing model from the issue text.