Constrained Deep Hierarchical Reinforcement Learning with Differentiable Formal Specializations