Multiple predictions per observation for time series

As an evalml user, I want time series pipelines to make multiple predictions per observation. For example, using the data from the past week, predict the target values for the next three days.

How should this parameter be specified?
* As an extra parameter to pipeline `predict`
* At the pipeline init level

I think the answer depends on whether this parameter should affect how the pipeline is fit.

If we go with option 1, then the pipeline is fit to only make the "next" prediction (whether that's tomorrow or k days from now) and we can get predictions for the entire horizon by recursively calling `predict` on the predictions. I think this would work but if the pipeline isn't fit on its own predictions, the errors would compound towards the end of the horizon.

If we go with option 2, then we can either fit a different estimator for each timestep in the horizon or we can fit the estimator with a mixture of its own predictions and ground truth (in the hopes it will learn to correct its mistakes).

The answer will depend how this feature is integrated into AutoML.