Balancing Exploration and Exploitation in Hierarchical Reinforcement Learning via Latent Landmark Graphs