Created by: grwlf
Change from this PR allows nn_trainer to save model on existing from train function correclty. We assume that model saving shouldn't be ignored if its parameters were changed since the time of previous save. Since chainer.train_on_batch changes parameters, we need to mark the model as 'dirty'.
Previous version of DP didn't clear _saved flag at all. As a result, the trainer didn't save the last version of model if its save method was called at some point during or before training, e.g. in _validate_first mode.