Scoring efficiency

[sbfnk]

In our weekly evaluation we keep re-scoring the whole data set of truth and forecasts, including values from 2 years ago. This is causing memory issues in the github action (temporarily fixed by running locally). It would be good to think of a system to only score the most recent forecasts, especially now that hospitalisation data are treated as final after 28 days (and perhaps so should the other data).

Currently responsible for #2635.