Here is a quick summary of the results I'm getting with the state of this PR on the HVRP benchmarks I gathered.

Comments:

1. The gaps are quite hard to get down, those results have involved quite a lot of testing and introducing a new operator. The HVRP is significantly harder than the CVRP since there is an additional affectation problem to decide which vehicles to use. I've tested many things around this in the scope of our solving framework but there is no simple rule that works across all instances to decide on vehicles to pick based on fixed/variable cost and capacities.
2. I feel this is good enough for now because:

• as usual the real-life instances are probably much simpler usually (especially with a more or less adjusted fleet, the hard part with the benchmarks being that they have an "open-bar" setup on a lot of different vehicle types);
• computing times are still quite low (this is not even a release build) which leaves us with some room for further improvements later on.