Learning Mixture Models via Efficient High-dimensional Sparse Fourier Transforms