Issue1102

Currently, almost every lookup on a new-style instance (*Derived) does
in fact two lookups: first finds the '__getattribute__' slot in the type
mro, then the actual attribute lookup takes place. And, if
__getattribute__ raises AttributeError,  another mro lookup takes place,
this time for __getattr__.

This is somewhat solved on old-style instances (PyInstance), where the
__getattr__ slot is cached on PyClass. Although it has a problem: the
cache is not invalidated/refreshed when the __getattr__ slot is
dinamically changed.

We also should take into account that __bases__ can be dynamically
changed: http://bugs.python.org/issue635933

r6161/r6210 fixed PyClass's cached __xattr__ slots not being refreshed 
when changed

r6844 / r6845 totally avoid the __getattribute__ lookup/descriptor call 
by fast pathing if the type uses object.__getattribute__. This gives us 
a pretty good speedup, we beat CPython on richards now, @ about .9 
whereas r6843 is about 1.2/3x slower

CPython does an optimization here where it looks up the descriptor and 
fast paths if it's equiv to object.__getattribute__. Our's is based off 
PyPy's which is more agressive, we keeps a flag on the type to determine 
whether we should fast path, so that faster path also avoids the 
descriptor lookup