[ML] Explore if we can handle sudden changes in the period of seasonality better

We had a question which was due to a time series which appeared to well described as seasonal but with a HMM determining the period of the repeat. We don't really support sudden changes in the period of repeat. Any sort of Viterbi algorithm to determine how to decompose such a time series is not well suited to running online and so not suitable for our needs.

Although we allow small changes in period of a seasonal component over time, this is a drift process and doesn't adapt fast enough to such changes. We also continually test how best to decompose a sliding window, but this doesn't handle flipping backward and forward between different states well. Basically, the modelling is generally poor unless the changes are smallish and continuous or infrequent.

We do have one potential mechanism for dealing with this sort of thing slightly better: change detection. In the problem case, it did appear that the signal was literally a time scaled copy of itself in different states. Perhaps we could add a new change type which checks for such events. We would likely still want to wait at least one period, and probably more, before we could be confident that such a change had occurred, particularly since a priori it would be unlikely.