This saves both time and memory. Before, data set was preprocessed
separately for each number of clusters and then also when computing
silhouettes. If used with from-to this made many in-memory copies of
data.

Also compute silhouette within worker threads. Fixes bug introduced
in Orange 3.22, when Orange started also computing full silhouette
scores for big datasets). Furthermore, they were computed in the main
thread, which blocked the UI.

Because silhouette scores are computed in worker threads,
__preprocessed_data attribute is not needed anymore.