Adaptive Gradient Methods at the Edge of Stability