ie the way that __getattr__ ends up getting called.  Two main optimizations:
- switch the initial attribute access (ie checking to see the attribute exists
  before calling __getattr__) to use a non-throwing api.  For the typical case,
  even though the throwing would be a C API throw, the construction of the
  AttributeError is relatively expensive, and the object would be immediately
  discarded anyway.
- add rewriting to the function

They both roughly cut out half the overhead of accessing attributes on
classes with __getattr__.