A multimodal and temporal foundation model for virtual patient representations at healthcare system scale — ThinkLLM