implement predict. generic for all models

[wwiecek]

To bring this in line with other rstan packages.

[wwiecek]

posterior_predict could have a 3-fold function:

• alias for effect_draw (average effect in a new setting)
• predict distributions for new studies with fixed characteristics (newdata = ...)
• predictive distributions for the existing K studies which is sort of implemented in predict. generic, but should use group_effects instead

[wwiecek]

What if newdata has logical type columns but either FALSE or TRUE is not present? Will model.matrix not break things?

[wwiecek]

Update: as of 0.7.1 we only need to implement the third bulletpoint in the list above.

[wwiecek]

Also need posterior_predict() and posterior_interval(). See #34 for much older conversations on this