Improve the rebuild of user activity points

Improve the rebuild of user activity points

The previous implementation relied on `LIMIT … OFFSET …` which has an incredibly poor performance on large data sets. Changing this to `BETWEEN` yields a consistently low latency.

In the previous implementation, up to 92% of the request was spent fetching the target rows.

For a test with 500k records this cut down the time to rebuild from 2:30h down to 20m.