PipeOp that constructs a data-dependent Graph and runs it

[mb706]

The user gives a function as parameter that takes a task and returns a Graph. That graph is then trained on the training data and the trained Graph is used for prediction.

Some notes:

• probably not just a PipeOpTaskPreproc, but a whole PipeOp, with construction argument: number of input channels (default vararg), number of output channels (default 1)
• parameter function is called with do.call, so user doesn't need to bother with list access
• Difficulty here is that cloning gets a bit difficult maybe?