Convexifying Sparse Interpolation with Infinitely Wide Neural Networks: An Atomic Norm Approach