Cross-modal prototype learning for zero-shot handwritten character recognition