This fixes a rather central bug, which causes duplicated indexes if
edges are removed and then new ones are added. Edge indexes are now
recycled as they are removed and then new ones are added. This still
guarantees O(1) complexity when adding or removing edges.