Warn at fit time (instead of construction time) for GLMM with dispersion parameter.

[palday]

Since the major problem is the evaluation of the log likelihood and not PIRLS, evaluation of the linear predictor, etc., there may be valid reasons for constructing a model from scratch and using it for simulation, prediction, etc.

This would also make it somewhat easier to construct test examples since we could just use simulate instead of writing out a simulation ourselves.