Adapting the Linearised Laplace Model Evidence for Modern Deep Learning