pickup and delivery option with cspy

Just had a quick google about the pickup and delivery and found this paper. End of page 10 basically says that the bidirectional algorithm won’t work due to dominance conditions for backward labels. Might still be worth running these with direction="forward" after we figure out how to model it

_Originally posted by @torressa in https://github.com/Kuifje02/vrpy/issues/14#issuecomment-622475443_

[torressa]

Had another go at this. I’m playing around with allowing non-elementary routes (will submit a PR soon) and it works much better (at least better than the LP)!

The code for this (PDP) is very specific to VRPy, so not sure what the best approach would be here. We could move the VRPY-specific version of cspy out as a separate repo or maybe even here. With this we could go further into moving more subproblem parts into C++, like writing the REF functions (after passing the relevant data).

[torressa]

I’ve updated the implementation (on the cspy side) to what should be OK now (updating open requests sets on the fly, as opposed the previous terribly inefficient prototype). The optimal solution is reached much faster than the LP, however it takes a very long time to converge, making it actually slower.

Not sure how to proceed here, the paper suggests adding several inequalities (pickup-delivery related and others) to strengthen the formulation and probably also help with convergence. This maybe raises the need to implement the generic cutting framework so we can add all try out new and varied inequalities, maybe non-robust ones (which add resources to the subproblems).