[ML] Forecast upper bound less than lower bound

[tveasey]

This is an error and we should as a minimum take steps to ensure it doesn't happen: there is currently no error checking for this case. It would also be good to understand how this can happen (reviewing the code doesn't show up any obvious edge case which could lead to this).

[tveasey]

I haven't been able to reproduce this yet. One possibility would CTimeSeriesDecomposition::smooth. This could in theory cause bounds to get inverted if the change in the upper and lower bounds across the point where we join the models is very different. I think the pragmatic thing to do is to add in some edge case handling in the forecasting code for 7.9 to explicitly stop this occurring and continue to try and understand the underlying cause.