Lovely question! Check out the explanation to algorithm 4 in the paper, I'm not sure I explained it well enough in the writeup.

You are right, nodes are only capable of opening channels between themselves and someone else, so this algorithm would have to be adapted. The algorithm selects 4 nodes according to some properties. Then the easiest modification would be to open channels with them to a fifth node (your node).

Another idea is to just insert yourself as one of the 4 nodes. You could do this artificially (replace one of the four with yourself at random), or you could exclude candidate nodes from the algorithm. By excluding nodes with a greater degree than you, you would become node i in the algorithm. Then just keep the actual channels you can open.

Robustness tells us how well a network can maintain service when nodes become unavailable (failure/attack/liquidity). Using the method from the paper should distribute closeness centrality and improve path diversity between nodes. But you're right, I don't think people are necessarily incentivized to open channels in this way, it's more of an altruistic strategy.

It would be interesting to see how well the robustness of LN could be improved by a limited number of rewirings.