Require >= 4 images in dataset

## Context
During testing with a dataset of 2 images, @derekthirstrup ran into a bug during training. Effectively, training would fail because cyto-dl did not like that our `test.csv` and `val.csv` files were empty. Then, because training would fail, no `metrics.csv` file would be created. Since we haven't addressed #438 , this would cause a different error: file not found (since we look for final loss when we display the training finished dialog).

## Solution
After discussing with @benjijamorris , we decided that we should require a dataset of >= 4 images since datasets smaller than this are unlikely to produce a reasonable model anyways.