Variational Dynamic for Self-Supervised Exploration in Deep Reinforcement Learning