Resolve TROVE-2017-006: Regression in guard family avoidance in 0.3.0 series

In 0.3.0, when I revised the guard selection code, I got the "don't use your exit as first-hop if even if it's your guard" logic right, but I accidentally omitted "don't use anything in the exit's family as a guard."

The impact here is that some circuits will still use a guard on the same circuit as an exit even when the two are in the same family.

(Putting this issue up on the tracker now because knowing about it does not (much) help an attacker exploit it.)

This is TROVE-2017-006 and CVE-2017-0377.