consider "cluster lag" for use in transition models

[knaaptime]

as described here

[knaaptime]

after testing this out a bit yesterday, I think this is still an interesting idea, but the linked implementation is wrong

In that link, I create a cluster model, then take the lags of the input variables and use those to create a different cluster model

instead, we need to

1. build a cluster model
2. take the lags of the input variables and scale them using the same scaler as in the original model
3. predict cluster membership based on the fitted model from 1

but this creates some complexity. We dont need the "cluster_lag" until we're doing transition models, but we need the original clustering instance (and scaler instance)to create it. We could add the scaler (or just the pre-computed cluster lags) to the ModelResults class so we have it on hand