cpu-gpu compatibility

[rasmuspjohansson]

Evaluating models usually goes faster with GPU enabled. This simple change enables model_weights_as_vector() and model_weights_as_dict() to work also when torch use GPU for computations.