B-cosification: Transforming Deep Neural Networks to be Inherently Interpretable