MADA: Meta-Adaptive Optimizers through hyper-gradient Descent