TopoReformer: Mitigating Adversarial Attacks Using Topological Purification in OCR Models