Zero-Shot Character Identification and Speaker Prediction in Comics via Iterative Multimodal Fusion