Posterior Coreset Construction with Kernelized Stein Discrepancy for Model-Based Reinforcement Learning