Efficient Weight-Space Laplace-Gaussian Filtering and Smoothing for Sequential Deep Learning