Rank traversal bugfixes

[faithokamoto]

Changelog Entry

To be copied to the draft changelog by merger:

• Better distance indexing in complex DAG snarls. Distance indexes should be re-made.

Description

Three bugfixes, which were found in decreasing order of complexity. These all affect the traversal which assigns ranks to children of snarls.

1. Actually implement a proper DAG traversal algorithm, as was written in comments, by adding tips into the list of sources that should be traversed from.
2. Fix the tip-finding algorithm to properly handle chains which function as tips. In particular, use the correct chain bound node, and its corresponding orientation, when checking for edges incident to a chain side. The previous algorithm checked both orientations from the chain start node, instead of checking from start/end.
3. Shift raw indexes for all_children into ranks, by subtracting two, before checking whether that rank has already been visited. This was done everywhere except one spot.

I also slightly modified the way we stored tips and bumped the distance index version vgteam/libbdsg#238 with the latter being because we really want people to re-index to get better distances. Old indexes are technically compatible but will get worse chaining mode alignments.

[faithokamoto]

bugfix no 3 doesn't have a test because it turns out to be difficult to configure a situation where the initially-assigned ranks (arbitrary order in the children vector) have exactly the relationship needed.

[faithokamoto]

Results came back; slight speed drop and slight increase in accuracy. About 0.5% of reads which used to be mapping incorrectly will now map correctly.