DirectPipelineRunner: support for unbounded collections

DirectPipelineRunner currently runs over bounded PCollections only, and implements only a portion of the Beam Model.

We should improve it to faithfully implement the full Beam Model, such as add ability to run over unbounded PCollections, and better resemble execution model in a distributed system.

This further enables features such as a testing source which may simulate late data and test triggers in the pipeline. Finally, we may want to expose an option to select between "debug" (single threaded), "chaos monkey" (test as many model requirements as possible), and "performance" (multi-threaded).

more testing (chaos monkey)

Once this is done, we should update this StackOverflow question:
http://stackoverflow.com/questions/35350113/testing-triggers-with-processing-time/35401426#35401426