Spike - Compute rolling windows features for time series

[freddyaboulton]

The outcome of this issue is a design document detailing how we can augment the existing set of time series transformers with new time series transformations.

I'd like us to prioritize leveraging ts_fresh because:

1. These primitives are time-aware and respect series indicators - quickstart
2. Support forecasting with rolling transformations - demo
3. There's an integration with sktime - docs

[freddyaboulton]

I just filed #2733 which is related to this.

Another related thought is to make max_delay an optional parameter to increase the amount of data that can be used to generate features.

[freddyaboulton]

Progress so far:

We've been chatting with the featuretools team about adding rolling window transform primitives to featuretools. Current design is here

[freddyaboulton]

As discussed at our synch, we don't need to wait for featuretools to add this functionality to close out this issue.

We can make progress by writing a simple rolling average component in base pandas that we can then swap out for the more mature featuretools implementation.

To close out this issue we should therefore:

• Write a design document detailing how to add a component that will compute a single rolling average feature. This document should detail how we can eventually swap out the logic for the featuretools primitives talked about in the design document above.
• Implement that design document.