Counterfactual (Non-)identifiability of Learned Structural Causal Models