Understanding the Initial Condensation of Convolutional Neural Networks