circmgr: Don't count channel negotiation towards circuit built time.
This is important, since some circuits require us to negotiate a new channel, and some don't: by adding this variability, we can easily learn a timeout based on pre-existing channels to our guards, but then later time out if it takes a while to open a TLS connection.
This is a possible solution to part of #2079, a bug where we "learn" a circuit timeout that is too low for us to satisfy.
I don't think that this branch will be a 100% solution: IIUC our retry logic should have reset the timeout, or marked all the guards up (or something like that) when it found that all the guards were down and all the connections were failing.