ECCO: Can We Improve Model-Generated Code Efficiency Without Sacrificing Functional Correctness?