Self-supervised Sequential Information Bottleneck for Robust Exploration in Deep Reinforcement Learning