This gives us considerably better estimates when we don't get the exact
length from the backend.

Hafas even has separate length values for each line segment along a path,
but those values match our computation within a 0.1% margin, so not worth
explicitly storing that.