Leveraging large language models for the deidentification and temporal normalization of sensitive health information in electronic health records