Fine-Grained Causal Dynamics Learning with Quantization for Improving Robustness in Reinforcement Learning