Boosting Robustness in Preference-Based Reinforcement Learning with Dynamic Sparsity