High-dimensional Asymptotics of Feature Learning: How One Gradient Step Improves the Representation