Streaming train tweaks

This adds a few more log statements to help clarify how decisions to re-train or re-calibrate are made online.

This also adds logic to start of non-incremental classifiers with some amount of initial lookback when they first begin training. This is done to avoid as much of the burn-in period as possible when we re-start the pipeline. We also note that this is only needed for the non-incremental training jobs because everything else acts incrementally. Incremental training jobs (duh) pick up previous data when they read in the preferred model before the main loop. Preferred CalibrationMaps, similiarly, are read in before the main loop and act incrementally. Therefore, it is only the non-incremental classifiers within the training jobs that need special treatment.