$predict should throw an error if PO is not trained

[mb706]

Otherwise we get weird error messages when state is unexpectedly NULL.

Weird that we did not have this already.... maybe check if we do have it somewhere else and it does not apply to all pipeops, for some reason? (e.g. we only check in Graph?).

Example:

po("imputeoor")$predict(list(tsk("iris")))
#> Error: Input task during prediction of imputeoor does not match input task during training.

may be a minor breaking change if someone unsed "stateless" POs without training them, before. But this should really not have been allowed to begin with.