A Fenchel-Young Loss Approach to Data-Driven Inverse Optimization