Maximum Entropy Gain Exploration for Long Horizon Multi-goal Reinforcement Learning