Cleaned up LamarrPropagator; Residual-style generator logic moved to the imported shared object

The tracking GANs are designed to learn deviation of the target distribution from a standard multivariate Gaussian. This is reasonable because of the quantile preprocessing. The logic implementing this "deviation" was in LamarrPropagator before, but this is suboptimal because in a future modification of the model this "residual-like" logic might be removed. I moved the necessary steps out from Lamarr and included them in the pluggable model.

Note: this will break Lamarr as long as the parametrizations are not updated.