We should update the parent and update the g value for *that* node, and add a
comment that g values can be too large if we don't reopen closed nodes (because
they are the g values at the time the node was generated). Sounds better than
producing a longer plan than necessary.

I noticed that for some states, the g-value and and path found using trace_path()
are not consistent, when using eager GBFS.
This does not happen when I remove the update_parent() call, so I commented it
out for now.