Sample-Efficient Learning of Correlated Equilibria in Extensive-Form Games