[Python-3000] Armin's attribute lookup caching for 3.0

[Neil Toronto]

I couldn't help myself. Such a beautiful thing had to be spread, and it 
was *easy* to put it in 3.0 because types are simpler. A patch is here:

     http://bugs.python.org/issue1568

If users are going to be encouraged to subclass from the ABC hierarchy 
for new container and numeric types, they'll likely want something like 
this. Both pybench and pystones are faster (pystones because its classes 
are now instances of type), and those exercise attribute lookups on 
classes with MROs no larger than 2. The pybench scores are good in 
general (with all lookups very good), except SpecialClassAttribute, 
which spends half its time doing class attribute assignments. Another 
surprise is TryRaiseExcept - why should that be faster?

I found updating caches from setattr to be faster than invalidating 
entries. (It could easily go back to invalidating all entries for a type 
if that's preferred.) Adding extra TPFLAGS was unnecessary. It still 
assumes attribute names are interned (it doesn't have to but it's a 
little faster this way), but AFAIK this could only affect computed names.

Neil