Towards Generalizable Pathology Reports via a Multimodal LLM with the Multicenter In-Context Learning