On Tradeoffs in Learning-Augmented Algorithms