ENH: add_constant even if already there

[josef-pkt]

see http://stackoverflow.com/questions/32569835/valueerror-when-fitting-a-model/32589641#32589641

add_constant doesn't add a constant if there is already a constant present. However, when using it to create a exog for predict, then we still need a constant if the exog in fit has one.

proposal: add option to turn of the non-adding of a constant.

[josef-pkt]

This should be fixed by #3020