[Feature] Consider publishing a normalised pre-trained model

[alessandrofelder]

Is your feature request related to a problem? Please describe.
The cellfinder pre-trained does not come with statistics, which is expected to worsen classification in some cases, and increase amount of fine-tuning needed.

Describe the solution you'd like
Create a new pre-trained model based on the same training data as the current pre-trained model, but with the normalisation enabled.

Additional context
Follow-on from #588

[adamltyson]

Also linked to #450 (comment)

[matham]

The best way to get the required statistics for the existing training set without re-exporting all the cuboids is to do the following.

For each dataset in the original training dataset (i.e. each yaml/folder in the training set):

1. Open the original tiff (not extracted cubes) from which these cubes came from.
2. Label one cell (and one non-cell if curation needs that).
3. Save the cubes.
4. This will have generated a yaml file for this dataset. Open the yaml, copy the attributes with the statistics and paste it into the old yaml files for the cells and non-cells that came from this dataset.

[adamltyson]

Open the original tiff (not extracted cubes) from which these cubes came from.

I don't think we have these unfortunately.