On Theoretical Limits of Learning with Label Differential Privacy