Gathering algorithm statistics

[jsjason]

In dolphin-dnn of the open-source version dolphin, accuracy stats for the training dataset and validation dataset are being gathered at the parameter server alongside the training model. This not only makes the application code complicated, but it also kind of misuses the CPU cycles of the server, since gathering and reporting such stats are not vital parts of the algorithm itself. It'd be better if we do the gathering elsewhere for dolphin-async, maybe from the Driver.