add spatial cleaning to inference workflow as postprocessing step

[jdegerickx]

in the demo notebook I implemented a simple majority vote based on what has been done in HRL, but this still requires further testing and tuning. In HRL in the end we used a more advanced technique based on gdal sieve --> investigate whether this is possible and desirable here...

[jdegerickx]

@kvantricht, not sure who to assign this to?

[kvantricht]

Options:

• implement on worldcereal-classification side (short-term solution), by integrating postprocessing into the classification UDF (saving postprocessed label, max probability and the original class probabilities)
• implement a postprocessor method/UDF within openeo-gfmap which is configurable by other libraries making use of this method.

@kvantricht to setup meeting with @HansVRP and @GriffinBabe to discuss option 2 @GriffinBabe to link here to some existing code from HRL as a reference

One workflow that needs to be part of this is based on gdal_sieve. Question: is such operation available within OpenEO? To be checked with @jdries

[GriffinBabe]

The smooth and reclassify operation, which I propose to start with is implemented here: https://git.vito.be/projects/APPL/repos/cropclass/browse/src/cropclass/postprocessing/operations.py#93µ