make single pipeop application easier

[mb706]

p = po("encode")
task_preproc = p$train(list(task))[[1]]

[mb706]

potential problem: we don't want to tempt people into doing the wrong kind of preprocessing on train and test set

[mb706]

suggestion (bb):

outdata = preproc("encode", indata, paramlist, state = state)

[mb706]

maybe we want to have preproc(po("encode", paramlist)) instead, more consistent and would also work with PipeOps that are not in po().

[mb706]

could also overload the %>>% operator here.

[mb706]

maybe: whenever state is given, or when optional var predict = TRUE (default FALSE) do prediction, otherwise training. We live with the fact that the state gets saved to the 1st arg pipeop. No reason not to make this work with graphs as well.