base BCE weighted on tokenized data

[sfluegel05]

Previously, the class-based weighting used the pkl-files from processed-main (untokenized). This has two disadvantages:

1. If some instances get thrown out due to not being tokenizable, the weighting will be a bit off
2. If a user wants to provide their own tokenized datasets, they don't have to change the untokenized one as well

I also added a safeguard to avoid concatenating empty prediction lists which would raise an error.