Add feature to train on predicting next n time steps

[ali-ramadhan]

Instead of training on data pairs u_n and u_n+1, we can try training on u_n and (u_n+1, u_n+2, u_n+3) for example.

[ali-ramadhan]

Closing as it's kinda there in PR #4 but I think we want to train on the entire time series at once.