Deep reinforcement learning uncovers processes for separating azeotropic mixtures without prior knowledge