Non-Uniform sampling / Restriction of infill-opt-domain

[mb706]

E.g. my optimization happens over a simplex. One possibility is to map from the hypercube to the simplex using a trafo, but that would mean the surrogate model fits the performance over hypercube, which could give a poorer fit than if we could fit the model to the performance over simplex coordinates directly. An idea would be to have a transformation that happens between sampling/optimization and model fitting.

[sumny]

related to #17