The auto-regressive model allows for different window shapes
and different lag sizes.

Although most likely to be used as a reference for modeling
noise in AV1, the model is currently parameterized more generally
than AV1 needs.

I will add an example (hopefully with a denoiser) in future
commits.

Change-Id: I1ba1067543601c2c01db4970d42766bb35da77f0