boxstacks: skip unsized bins in Solution::feasible_axle_weights

[HansBug]

Fixes #575.

SolutionBuilder::build calls update_indicators once per bin, and each call recomputes the aggregate feasibility over every bin. The per-group weight vectors of a bin are only sized by its own pass, so while the first bins are processed the later ones still have empty weight / weight_weighted_sum vectors, and feasible_axle_weights indexed them anyway: undefined behaviour that _GLIBCXX_ASSERTIONS reports as an out-of-range std::vector::operator[] and that is an access violation with MSVC. feasible_total_weight already skips such bins; this does the same. The last pass sees every bin sized, so the final feasibility is unchanged.

Reached by every multi-bin solution rebuilt on another instance (Reduction::unreduce_solution, on by default since 24decc4, and InstanceFlipper::unflip_solution) and by reading a multi-bin certificate.

Two commits:

• boxstacks: skip unsized bins in Solution::feasible_axle_weights — the guard, plus BoxStacks.SolutionBuilderTwoBinsBuiltAtOnce, the smallest solution that reaches it.
• boxstacks: certificate-reading tests for multi-bin solutions — a new instance data/boxstacks/tests/bin_packing_postal_cartons_eur_pallets (three postal carton sizes on 1200 x 800 x 1800 EUR pallets, bin packing, optimal in two pallets so optimize reaches the certificate without a time limit), added to BoxStacksOptimizeTest; and BoxStacksSolutionBuilderReadTest, which reads a multi-bin certificate through SolutionBuilder::read + build and checks bins, fullness and feasibility, for this instance and for bin_packing_two_item_types_pallets_time_limit.

Validation (Ubuntu 24.04, GCC 13, RelWithDebInfo with -fsanitize=address -D_GLIBCXX_ASSERTIONS, CLP + HiGHS):

• PackingSolver_boxstacks_test: 21 tests pass.
• With src/boxstacks/solution.cpp reverted to master and the tests kept: both BoxStacksSolutionBuilderReadTest cases, SolutionBuilderTwoBinsBuiltAtOnce and the postal-carton BoxStacksOptimizeTest case abort on the assertion in feasible_axle_weights.
• Windows (windows-2022, MSVC 19.44, RelWithDebInfo, HiGHS): the bin_packing_two_item_types_pallets_time_limit CLI run crashes with 0xC0000005 in feasible_axle_weights (solution.cpp:140) on master and completes with the patch (32 / 32 items in 2 of 4 pallets, 0.15 s); backtraces in the gist linked from [boxstacks] SolutionBuilder::build on a multi-bin solution reads past the empty weight vectors of the bins not processed yet (access violation on Windows since instance reduction) #575.

---

Update: added a third commit, boxstacks: accumulate axle weight feasibility per bin instead of rescanning, replacing the "skip unsized bins" rescan from the first commit. update_indicators(bin_pos) is only ever called once per bin_pos, in increasing order, and bins are never removed once added, so a bin's own axle weight feasibility, once computed, cannot change. feasible_axle_weights() is now split into a per-bin feasible_axle_weights(bin_pos) and an axle_weights_feasible_ flag ANDed with it right after bin_pos is sized in update_indicators, instead of rescanning every bin (sized or not) on every call. The no-arg feasible_axle_weights() is kept as a full loop over sized bins for existing external callers. Tests from the second commit are unchanged and still pass (21/21, PackingSolver_boxstacks_test, Ubuntu 24.04 GCC 13 RelWithDebInfo).