Improve parsing speed with profile-directed code tuning

Hi! I took a quick look at how Stem parses descriptors, and tried running cProfile over a networkstatus parse. I don't think I have all of the issues solved, but I managed to get a few of the hot spots calmed down for a speed reduction of about 30%.

I don't promise that these are all correct, though I am pretty sure that the first one is a big win that you should take. It accounts for most of the speedup that I'm seeing.

Thanks again for the patches!