Remove classes from interpolator training, if they are underpopulated.

I think, the best way would be to remove interpolation classes from the interpolator training, if they are underpopulated.
I think, the exact number of needed systems (probably even within a cluster) will depend on the interpolation method. Hence, I think we should add this to the interpolation class itself. E.g. when the normalization is done, we should know the number of systems participating there already, thus I'd add there to mark an interpolation class there as to be excluded from thereon.
Ideally, we would not discard those systems, but move them to another interpolation class, but this might be very difficult to decide and may cause the interpolation errors to increase significantly. Therefore, an easy way is to exclude them in the first place.
In case, we'd still train the classifiers with this interpolation class, we can add a fallback for classes excluded from interpolation training to default to nearest neighbor, which should always work.