Faithful and Efficient Explanations for Neural Networks via Neural Tangent Kernel Surrogate Models