Completely rewrite the loop-detection algorithm used to check game

Completely rewrite the loop-detection algorithm used to check game
completion, _again_. In r6174 I changed it from dsf to conventional
graph theory so that it could actually highlight loops as opposed to
just discovering that one existed. Unfortunately, yesterday I
discovered a fundamental graph-theoretic error in the latter
algorithm: if you had two entirely separate loops connected by a
single path, the path would be highlighted as well as the loops.

Therefore, I've reverted to the original dsf technique, combined
with a subsequent pass to trace around each loop discovered. This
version seems to do a better job of only highlighting the actual
loops.

git-svn-id: svn://svn.tartarus.org/sgt/puzzles@6283 cda61777-01e9-0310-a592-d414129be87e