When calling `merge_chain` from `interclexp`, this code apparently did not
anticipate that both the condition `ND_rank(agtail(e)) == ND_rank(aghead(e))`
and `ED_to_virt(prev) != NULL` could be true at once. In this case, the merge
can happen but the `.to_virt` member of `e` needs to be overwritten; it is not
`NULL` on entry to `merge_chain`.

This issue appears to have existed since the first revision of Graphviz.

These is a lot of uncertainty around what the expected behavior of this code is
and whether this fix is correct. So here is some of my background thoughts:¹

  `merge_chain` has a leading `assert(ED_to_virt(e) == NULL)`, presumably trying
  to express the assumption that it is not overwriting an existing back pointer.
  But the code here in `interclexp` seems to be deliberately trying to overwrite
  this pointer. There seemed to me two possible local fixes:

    1. Remove the assertion
    2. Write `NULL` to `ED_to_virt(e)` in advance

  The first possibility seemed more risky. This would allow other functions to
  accidentally call into merge_chain with a non-null `ED_to_virt(e)` and for
  this to go undetected.

  On the other hand, it's possible my change here is mistaken and what the
  original author intended was for the true case of the branch on line 181 to
  continue after updating `ED_to_virt(e) = prev`, making the call to
  `merge_chain` not reached in this scenario. It certainly looks pretty weird
  for the code after my changes to set `ED_to_virt(e)` on lines 181-184 and then
  subsequently overwrite this value on line 187. But I was guessing the initial
  write is relevant for the case where we continue.

  As you can tell, all of this is (barely) educated guess work. I find the
  intent of this code very hard to determine.

Gitlab: fixes #121

¹ Quoted from discussion on
  graphviz/graphviz!2724 (comment 1005393861).