[ENH] Extend forecasting benchmark to global setting

[benHeid]

As far as I know, currently the forecasting benchmarking (as well as evaluate) is not working for global setting. I.e., fitting a forecasting model on time series 1:n and evaluating it on n+1:m with n<m.

@fkiraly please mention if it is already implemented.

[fkiraly]

No, not implemented. For this, evaluate would have to be adapted I think.

I would be surprised if it worked out of the box - it could be in theory because there is an InstanceSplitter which splits a collection of time series by instance, so it might be if we pass that it works. Even if, this is perhaps not the best user experience.

Possible improvements:

• if sklearn splitter is passed, it should be wrapped in the InstanceSplitter
• for exogenous data, we need to find a mechanism that the same splits are applied, if the y-splitter is InstanceSplitter or has the "instance-wise" tag.