Implement time-based windowing

[tk-king]

• Create a custom class "window" which supports time-series-parts with different lengths

[omeryagmurlu]

Apparently, the dtype=object approach was working without any problems, and I had made a mistake during implementation. We don't need any custom classes for window.

We still need to figure out how to decouple (or couple) windowing and feature extraction though, NoFeatureExtractor still doesn't work with varying window sizes, and no custom window class would help because sklearn doesn't work with varying length features anyway.