Allow the user to ONLY calculate aggregates while reading in an existing utterance-level dataframe

[xehu]

Currently, when the FeatureBuilder is run, we go through the entire process --- generating utterance-level features from scratch, then conversation-level features, then conversation-/speaker-level aggregates. There might be a world in which a user wants to pass in an already-featurized utterance-level dataset, and they simply want the aggregated features. This would potentially allow the FB to run faster, without going through the whole featurization process again.

Major issue: how would we validate the utterance-level data the user passes in? What if they pass in something that's junk? Right now, since we control the entire pipeline, we can guarantee that the utterance-level features are legitimate. If we allow the user to pass in their own version, we need to think about how we want to validate it --- otherwise, the aggregates are just garbage-in, garbage-out.

Another variation of this: what if the user wants only the features at the conversation-level, and they want to skip the generation of the chat-level features?

[xehu]

This is also related to: #285, which proposes that users append features to a dataframe of previously-generated features. That proposal gets into the same issue with data validation (how can we be sure that the supposedly previously-generated features is legitimate?), so we need to address these two together.