From Logistic Regression to the Perceptron Algorithm: Exploring Gradient Descent with Large Step Sizes