Fully probabilistic deep models for forward and inverse problems in parametric PDEs