I think it helps to distinguish between training metrics and evaluation metrics, and between global training metrics and local training metrics. When we talk about evaluation metrics, as @AlvaroFuentes said, a loss function can always be defined for decision trees, in the same way as for any other model. In training, it is true that often a global metric is chosen and training attempts to optimize over that metric *. But training does not have to be this way, and in the case of decision trees, training proceeds through a greedy search, each step based on a local metric (eg, information gain or Gini index). In fact, even when a global training metric is defined (say, likelihood), each step in training is still is evaluated based on some local metric (say, gradient of likelihood) and hence in a sense "greedy"; it is just that in this case the local metric is inspired by the global one. And in both cases there is no guarantee that a greedy search based on some local metric actually optimizes the global metric (eg, local optima).
* Side note: This training metric is often different from the evaluation metrics, chosen for its nicer mathematical properties to help the training; eg, likelihood, L2 or cross entropy vs accuracy or AUC.