Uncertainty-aware transfer across tasks using hybrid model-based successor feature reinforcement learning