ENH: implement __iter__ for window objects

[jreback]

xref #11603

might be slightly non-trivial as the current impl does NOT use an explict iterator, rather does a sliding window in cython and marginal calcs.

[kawochen]

I don't know if this is too ambitious, but for custom rolling aggregation, if you can let users supply a StateVariables object which is initialized using the first window, and supports two methods, evolve() and compute(), then you may be able to retain the efficiency when the evolution of the state variables is governed by a simple set of equations. __iter__ becomes a special case of this where the StateVariables contains the whole window, and evolve moves the window forward/replaces the old window with the new window, and compute just returns the data structure inside.

[selik]

Currently, iterating over a pandas.core.window.Rolling object will produce bizarre behavior. For DataFrames as it looks for columns named 0, 1, 2, etc. until it hits a KeyError. For Series it raises AttributeError as it looks for a "columns" attribute.

Until the __iter__ is implemented correctly, perhaps there should be a temporary measure of simply making it non-iterable by raising NotImplementedError.

[jreback]

@selik if you want to submit a PR to raise NotImplementedError would be great!