Efficient and Optimal Learning of Discrete Distributions with Person-Level Privacy