Implement a sample single cells command

[jccaicedo]

We want to be able to generate fixed training sets of single cells by cropping and storing the single cells in disk. In that way, the training procedure can be simplified and the queueing and workers can be dropped all together, also paving the way for multi-gpu training.

The idea is to run the current cropping workflow once without augmentations, and instead of adding the example images to a queue, we should save the resulting crops to disk with an appropriate naming convention.

[jccaicedo]

Open questions:

• What would be an appropriate naming convention?
• Does this require a separate metadata file? (hopefully not).

[jccaicedo]

The functionality has been implemented here da0a993 and is currently discussed in the PR #238

In short, single cells can be efficiently cropped to create a static dataset and a new crop generator allows us to use that dataset for faster and potentially more accurate training.