Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models