Raise Exception when `GammaGammaModel` does not converge properly

[ColtAllen]

If a fitted GammaGammaModel has a q parameter less than 1, all predicted spend values will be negative. This usually happens when the model is fit to heavy-tailed monetary value arrays with values in the millions.

We should add a check for this raising the following exception:

ValueError: Model did not converge properly. Scale monetary_value array by maximum value and refit model.

[ricardoV94]

I don't think we should raise after fitting. It's fine to have q < 1, except some expectation methods fail then (but not distributional methods for instance). We should warn/raise only on the specific methods where q < 1 is problematic.