Neural Tangent Kernels and Fisher Information Matrices for Simple ReLU Networks with Random Hidden Weights