Recurrent Neural-Linear Posterior Sampling for Non-Stationary Contextual Bandits