Multiple Optimality Guarantees in Learning: Computation, Communication, and Privacy as Constraints on Estimation