Last analysis using backward finite difference rather than centered

For now, AR3 analysis at time t depends on t - dt and t + dt data. This is a problem when running diagnostic starting from the last analysis, when the forecast is starting, as it is biased.