Rethink RollingWindow abstraction

[arvoelke]

Some current issues with the abstraction:

• Provide public function(s) for pade delay polynomials #160.
• Difficult to provide a temporal signal as training data.
• Effect of radius is not clear (in particular its interplay with other methods/parameters).
• process parameter is too magical (in particular how it can change the way things are represented and scaled).
• Difficult to probe the postsynaptic current representation (i.e., the state for the spiking case); see LinearNetwork support for compensating for input/output synapse #125 (comment)