Model-Based Reinforcement Learning with Multinomial Logistic Function Approximation