[Feat] r.learn.ml2: Add support for semantic label

[veroandreo]

r.learn.train and r.learn.predict would benefit from having semantic label support as i.svm or i.maxlik in GRASS core. Having semantic labels support in these tools would allow the transfer of trained models to other imagery/raster groups with the same semantic labels. Currently, since map names are part of the models, the prediction only works if map names match.

[ninsbl]

See also: OSGeo/grass#3756

And here is a case where I used that function in a similar context as @veroandreo asks for here:
https://github.com/NVE/actinia_modules_nve/tree/main/src/imagery/i.pytorch/i.pytorch.predict

That may serve for inspiration...

There the concept is to have a JSON config/metadata file for a model that tells the module in which order to apply the different maps with semanic labels. It allows also to use maps with equal semantic labels (think a reference image). However, what is currently not that good implemented is a time series of more than two points in time maps as model input (think change detection in a time series of many points in time...

[ninsbl]

And this is how this is handled in i.svm: https://github.com/OSGeo/grass/blob/main/imagery/i.svm.predict/main.c#L155-L241 where the signature file defines the band order based on semantic labels...

[stevenpawley]

This would be a nice addition, should be fairly straightforward. I can try and take it on. The use of groups (but with the same raster names) was considered as the use case for transferring models to other images originally, but this was all before semantic labels, which is a nicer solution.