TODO

* Fix the startup benchmark.

* Lots and lots and lots more benchmarks. Some remaining ideas

    * ORM overhead compared to cursor.execute()
    * signal dispatch
    * more datastructures
    * more intensive url resolving and reversing
    * model validation
    * holistic request/response round-trip time (probably requires spawning
      a running server and testing it).
  
* Parameterized benchmark output: it'd be awesome if there could be, say, a
  single middleware benchmark that tested the overhead of each individual
  middleware component. This could be done with multiple benchmarks, but
  that starts getting into serious boilerplate. If a single benchmark could
  report multiple result sets, that'd rock.