[ML] Improvements to trend modelling and periodicity testing for forecasting

[tveasey]

Description
This is a merge of a feature branch for issue #5. This work has already been reviewed as individual changes; however, I had some merge collisions with other recent changes.

Effects
This affects results, both model bounds and anomaly detection, for all data sets for which we use trend and/or seasonality modelling. In practice this is most data sets for non-rare functions. We've investigated changes across our full benchmark set. In all but one case, where results have changed appreciably, our anomaly detection detection accuracy has improved. Of course the primary driver is forecasting, which is now more reliable, has greatly improved confidence intervals and doesn't display obvious pathologies, particularly for longer time ranges.

This introduces a mean 30% memory increase across our benchmark set. The reasons for this are unavoidable and can be broken down as:

1. The trend model itself is significantly larger (rather than 1 we now have 8 at different time scales).
2. We have the trend model all the time, previously this was created if we determined it was worthwhile to do so, but we need this model for forecasting and can't create it retrospectively.
3. The new style periodicity testing uses slightly more memory.
4. We have some extra state in the seasonal components for forecasting.
5. We sometimes find additional periodic components in the data (with associated accuracy benefits).

These costs can be offset against some recent memory wins going into 6.3.

[droberts195]

I noticed one thing that probably slipped between the cracks of the previous reviews of incremental changes.

[droberts195]

I don't think we should be doing this now the changed portion of the state is being upgraded.

Hopefully reverting this line won't mean regenerating any of the upgrade test files, but please can you check?

[tveasey]

Thanks @droberts195 it had indeed: corrected.

[droberts195]

LGTM