End-to-End Policy Learning of a Statistical Arbitrage Autoencoder Architecture