Multimodal Information Bottleneck for Deep Reinforcement Learning with Multiple Sensors