Remaining performance work to match C tor

In #88, we aim to get the largest bottlenecks out of Arti's client performance. With this ticket, we aim to implement "whatever is missing" out of the performance tricks from C tor, if those tricks will be worthwhile in improving Arti.

• Conflux, if implemented in Tor? (160h?) [needs ticket]
  • (needs breakdown)
• Circuit pre-building (40h?) [needs ticket]