In the Discover phase of UiPath Communications Mining, how does the system determine which groups of messages from meaningful clusters, and why is this process beneficial prior to supervised model training?