Make sure we take descriptors in consensus-descriptors.new into account as well

We did a bunch of investigations about the low amount of exits used for testing in #35 (closed). Over the weekend I hit this issue again, which bothered me sufficiently to look closer (there were several hundred relays less to test than in the consensus for some reason).

I finally think I've found the problem: First of all, we might exhibit a race condition here because exitmap is not waiting until all consensus descriptors are downloaded. But apart from that and which is more problematic, we never consider descriptors saved in consensus-descriptors.new. Doing that greatly improved things for me.