Investigate, implement and benchmark Vamana graph builder batch sizes reordering #419

        // Grow batches geometrically so early inserts can see each other in the sparse
        // graph; otherwise a batch collapses onto one hub whose pruning orphans nodes.
        // Holds batch sizes until the prefix sum below turns them into bounds.
        std::vector<size_t> batch_bounds;
        // Bound: ramp-up batches + full-size batches + leading 0.
        batch_bounds.reserve(std::bit_width(batchsize) + num_batches + 1);
        batch_bounds.push_back(0);
        for (size_t size = 1, inserted = 0; inserted < num_nodes;
             size = std::min(2 * size, batchsize)) {
            batch_bounds.push_back(std::min(num_nodes - inserted, size));
            inserted += batch_bounds.back();
        }
        // Only the final remainder batch can be smaller than its predecessor. Moving it
        // into the ramp keeps sizes non-decreasing.
        auto last = std::prev(batch_bounds.end());
        std::rotate(
            std::upper_bound(batch_bounds.begin(), last, *last), last, batch_bounds.end()
        );
        std::partial_sum(batch_bounds.begin(), batch_bounds.end(), batch_bounds.begin());