`predict_newdata_fast` that is actually fast in GraphLearner

Doing "fast" inference with `GraphLearner` is not really possible in principle, as long as PipeOps might want to operate on the actual `Task` objects. But we could have a subset of PipeOps, e.g. the `PipeOpTaskPreproc`, or only those `PipeOpTaskPreproc` that use `train_dt` / `predict_dt`, and make these work much faster. Maybe even more than that, because there are some `PipeOpTaskPreproc` that need complex Tasks during training, but not during prediction. For this, the `Graph` would need an alternative route where only `data.table` are passed to the `train_dt` / `predict_dt` directly (maybe we give the `PipeOp`s a `train_fast` / `predict_fast` route), and the `GraphLearner` would check whether the Graph can do this. Once we have this, `predict_newdata_fast` could actually be fast.