Feature: time series forecasting

[lsorber]

Ideas:

1. Fit a NeoLSSVM on y(t+1), then fit a new NeoLSSVM on y(t+2) | ŷ(t+1), and so on. This is more expensive than training a single t+1 model, but has the advantage that there's a direct estimate for the confidence interval of ŷ(t+k) with conformal prediction of the k'th model. Another advantage is that this method does not need 'future values of past covariates' as darts does.
2. Automatically determine the forecast horizon k by stopping when the confidence intervals are no longer meaningful.
3. Initially, only support past covariates and no future covariates. Later, add automatic future covariates based on the dataset's time series index using Time Axes Encoders. And after that maybe allow users to supply their own future covariates.
4. To keep things simple initially, don't allow supplying a new target to forecast a fitted model on (i.e. predict does not accept X or y). Instead, users can fit a new model if there's new data.