evaluation can receive metric values from anywhere

[AntonioCarta]

This PR adds a small feature to the EvaluationPlugin that allows to compute metrics outside MetricPlugins.
For example, let's assume that a strategy computes a loss function with many regularization terms, and we want to log each one separately. This is possible with the new API:

    # This is a strategy method, not a metric plugin.
    def criterion(self):
        if self.is_training:
            pred = self.teacher_targets.detach().max(dim=1)[1]
            loss_activation = -self.teacher_features.abs().mean()
            loss_ce = F.cross_entropy(self.teacher_targets, pred)
            loss = loss_ce * self.oh + loss_activation * self.a
            
           # Metric is logged here.
            xp = self.clock.train_iterations
            mval = MetricValue(origin=self, name='loss_ce', value=loss_ce.item(), x_plot=xp)
            self.evaluator.publish_metric_value(mval)
            mval = MetricValue(origin=self, name='loss_a', value=loss_activation.item(), x_plot=xp)
            self.evaluator.publish_metric_value(mval)

Doing the same with metric plugins is possible but much more verbose.