It previously crashed, since it called super(...).getattr(..) and the
first ancestor in its mro was bases.Proxy and bases.Proxy doesn't implement
the .getattr method. Closes issue #91.

In astroid.bases.Instance._wrap_attr we had a detection
code for properties, which basically inferred whatever
a property returned, passing the results up the stack,
to the igetattr() method. It handled only the builtin property
but the new patch also handles abc.abstractproperty.

This patch reverts some changes added by a couple of changesets, making
the context caching and scoping to work as before.
The changesets in question are:

 * 41b3bd589da0549ac061bc4c4b5d379cdbb1e10c
   Replace copy_context with some dynamic scoping.

 * 8d3c601
   Remove context.lookupname; make it an argument to infer() when appropriate.

 * partially reverts  048a42c.
   Fix some deep recursion problems.


There were some problems with these changes, which led to horrendous
performance when dealing with multiple inference paths of the same names,
as seen in these Pylint issues:

* https://bitbucket.org/logilab/pylint/issue/395/horrible-performance-related-to-inspect
* https://bitbucket.org/logilab/pylint/issue/465/pylint-hangs-when-using-inspectsignature
* https://bitbucket.org/logilab/pylint/issue/430/pylint-140-execution-time-and-memory

The reverted changes assumed that a context it's only passed to callees
and then destroyed, thus InferenceContext.push always returned another inference context,
with the updated inference path so far. This is wrong, since contexts are sometimes
reused, so the original context, the one before the .push call need to have the same
cache key in its path (this is actually what's happening in these mentioned issues,
the same object is inferred over and over again, but with different contexts).