Consider to scale taking descriptors but no consensus weight, as Torflow does

While working on #40051 (comment 2728407), I realized this is another thing that can make weights for the same relay diverge between sbws and Torflow.

In sbws, we decided (#33871 (comment 2519205)) to take the minimum of the descriptor bandwidth values and the consensus weight when the relay is unmeasured. Or only the consensus weight when there's no descriptor observed bandwidth (including burst), while Torflow is taking the max.

If we stop measuring relays without descriptor (#40076 (moved)), scaling would not be using only consensus weight, because there'd not be relays with no descriptor.

This will help to compare sbws and Torflow.

If we do, it should be done accepting a configuration/argument, so that we can go back to take these values correctly.