Transfer of Fully Convolutional Policy-Value Networks Between Games and Game Variants