The Mirage of Model Editing: Revisiting Evaluation in the Wild