A function to shrink a model

[K-RLange]

When training RollingLDA, the model grows in size in each consecutive time chunk. For large data sets this becomes problematic as all information regarding previous assignments and documents is still stored in the model, no matter how much time has passed.
It would be great to have a function to shrink the model by disregarding old documents and assignments in such cases to avoid the need for large RAM usage.