"Lossless" Compression of Deep Neural Networks: A High-dimensional Neural Tangent Kernel Approach