[CBM] Improve roll back logic efficiency

Problem

When merging, we iterate over the range files of all of the backups that have yet to be merged (here). We do this to prevent rollbacks (see MB-37092). This can be very inefficient, especially when lots of backups are being merged, as the number of range file reads is proportional to the square of the number of backups being merged.

Fix

Find a more efficient way of doing this while retaining rollback logic.