Latent Mamba-DQN: Improving Temporal Dependency Modeling in Deep Q-Learning via Selective State Summarization | Synapse