After a routine check on HVRP instances, I noticed the solution quality were highly degraded with this PR. This makes sense because the pruning argument based on geometric consideration somehow hold only if we assume homogeneous costs between routes.

In order to keep the benefit of the improvements here but avoid missing too many interesting moves, we should fine-tune and only prune route pair moves if the associated vehicles are "similar".

Turns out most of the HVRP instances (the DLP dataset) do not provide point locations... As a result the bbox checks where discarding all moves.

Nevertheless only pruning for homogeneous vehicle pairs is still a good idea since otherwise we're loosing too much quality.