Changes in the compiler and dictionary to make key sharing in classes more effective.

[markshannon]

Effective key-sharing is important for memory use and performance of class instances. With #72 it will become even more important.
Unfortunately, there are a few patterns that interfere with key sharing, when we generate the set of keys dynamically.
Anecdotally, in real-world applications there are many classes where keys are not shared.

E.g.

class C:
    foo = None #default value
    def __init__(self):
         self.bar = 0
    def meth(self):
        self.foo = 1

Building key sets dynamically will result in the set {'bar'}. Any objects that lazily update foo will be unable to share keys. By statically analyzing the class in the compiler, we can see that the set {'bar','foo'} would be a better choice.
Instances that do not define foo will waste a small amount of memory, but the overall saving from improved sharing would more than compensate.

To do:

• Gather data on the number and kinds of classes where key sharing does not currently work well.
• Discuss of Python Dev exactly when dictionary ordering should apply. Possibly write a PEP
• Implement generating key hints in compiler and using them in class creation.

[markshannon]

Closing this in favor of the more dynamic approach of #30 and #72