Right now, when constructing an INLJ, we only look at the first edge connecting the two subsets to find an index to satisfy the join. Now that we allow cycles in the join graph, there may be several such edges, and we should look at all of them to determine which index is best for the INLJ. 

When there are multiple index options, we should use the cost model to pick between these indexes.

When enumerating our index options, we can look at indexes that satisfy individual edges and indexes that satisfy all edges (or even some combinations of edges).