On the Empirical Neural Tangent Kernel of Standard Finite-Width Convolutional Neural Network Architectures