Fixes #570

While testing, I'm seeing situations where this PR currently hits:

This happens when custom costs and durations matrix are different. It means we have a faulty operator gain evaluation. I suspect this is a leftover from the costs/durations separation done a while back.

We're getting there but I'm still seeing weird things going on. For example when using a cost matrix that is the duration matrix doubled, then for some instances the solving path is changed at some point and we get a different solution.

AFAICT this is related to a problem in UnassignedExchange gain evaluation. There are probably still a couple values involving durations instead of costs.

After much digging, I think the last problems I'm seeing are not at all cost-related. My feeling is that comparing solutions across instances where we scale the costs is only a way to highlight a problem that is already present in the codebase, so I'll open a dedicated ticket for that.