Adaptive Rational Activations to Boost Deep Reinforcement Learning