Ignore failed sbws measurements as those might skew ratios/results and deal with missing desc_bw_obs_last

I was looking at a relay I knew had X observed (and, to make it easy, advertised) bandwidth. However, our dashboards were showing only X/2 as advertised_bw_avg significantly skewing our results. I think what ultimately happens is that we count measurement errors into our avg calculations. This is a thing we should probably filter out as both r_strm 0.0 and a potentially non-existing desc_bw_obs_last are skewing results.

Initially, I was thinking that the problem here was that measurements got done (successfully) before any desc_bw* were available. IIRC I have seen such cases. If that's possible and we'd take those cases into account as well then we'd have to filter them out, too, or switch to calculating the advertised_bw_avg directly from the server descriptor, which is definitely available at that point in time.