Understand how accurate the bandwidth authority estimates are

(Re-using text from Roger and Mike for this ticket description.)

It would be good to have a better understanding of how accurate the bandwidth authority estimates are. Why do some really fast relays get huge weights, and other really fast relays don't? Does it have to do with location of the measurers? What exactly is the trade-off between having fast nodes all nearby each other (and nearby the bandwidth authorities) in the network, and having nodes in geographically dispersed places?

We probably should figure out a way for the bandwidth authorities to utilize per-node as well as ambient circuit failure (legacy/trac#7023 (moved), legacy/trac#7037 (moved)).

There's a bunch of related stuff for TCP socket exhaustion, too. All of it probably involves some fairly diligent monitoring of results and experimentation though.