Add flexibility to predict() to allow random effects to be centred at value other than zero.

[nehill197]

This will give us flexibility to specify distributions for random effects that are not centred at 0.

E.g., gamma-distributed random effects are centred at 1. When predicting over a variable other than the random effect itself, the value is set to 0, which is not correct and causes issues when the gamma-distributed random effect is logged (log(0) = -Inf).

This can only be dealt with currently by having duplicate lines in the new expression with and without the random effect.

It would be ideal to have an argument to predict() that can override the default behaviour that sets the central value of the random effect distribution to 0.