Clean up connection timeout logic

In legacy/trac#6799 (moved), it was decided to keep TLS connections open for a random amount of time after they are idle, to defend against an attack that used internal Tor network connectivity information to determine Guard nodes (Slides: https://www.cryptolux.org/images/8/85/ESORICS-2012-Presentation-2.pdf Paper: https://eprint.iacr.org/2012/432.pdf).

Unfortunately, this logic (in connection_or_set_canonical()) is kind of a mess. Relays and clients are treated the same, and client connections are also kept alive for an additional hour by predictive circuit building even when otherwise idle, where as relays are not.

If we treat relays and clients separately for this timeout, we could reduce total client keep-alive time significantly (down to 30 minutes or so), and significantly increase relay connection lifetime. This should result in reduced total connection counts on relays, with better defenses against Torscan.

This is also needed in order to put reasonable bounds on padding overhead in legacy/trac#16861 (moved) for mobile clients. Furthermore, aside from the wieners running middle relays behind junky home routers who will whine about connection overflow, having a more well-connected Tor network is a good idea for many reasons (not just Torscan).