across packages

so we can use fit_predict! or fit_transform! instead of doing both 
separately

without an unnecessary one-hot encoding in the middle

Uses new form now and should work as intended

Deleted the old Krippendorff implementation from the main branch. A backup is still available in the Krippendorff-Archive branch

sorry for the squashed commit, this marks the version that was used to 
generate nearly all figures in the final master thesis

to better match how it is (likely) done in the prototype

this was a previous change that I just forgot to commit

because it previously treated high score = good (like accuracy), but 
high MAPE = bad

prio: would also take the whole block as one if multiple learnblocks 
were found  but only one was big enough. fixed now

bypasses model registration and just makes a PragmaticModel directly. 
Necessary when rebuilding a KB from a savefile

handle waste in construction and reconstruction instead of 
deconstruction. refactored nothingtuple shortcut to nothing

to match the versionthat is used by the autoencoder package thomas used 
+ some minor tweaks