Shape your Space: A Gaussian Mixture Regularization Approach to Deterministic Autoencoders