upgrading InfogainLoss layer: (1) incorporating Softmax layer to make the gradeint computation robust, much like SoftmaxWithLoss layer (see: http://stackoverflow.com/a/34917052/1714410 for more information). (2) supporting loss along axis