Representation Learning in Deep RL via Discrete Information Bottleneck