Support pickling tasks

[alcrene]

Although pickling simple minimal tasks sometimes works, this a) is untested and b) fails as soon as inputs or outputs involve dynamically created types (which are quite common in smttask). Thus we need

• Support for pickling tasks with dynamic types. These may be introduced via
  • task generators (task_generators.py)
  • multi-outputs (see base.TaskOutput.__init_subclass__)
• Comprehensive tests
  • Include all types of tasks (esp. all generated types) and both single- and multi-output tasks.
  • Test multiprocessing where:
    • Tasks are passed as arguments
    • Tasks are returned as results
    • Tasks are executed in the subprocess, and their result is returned.

The main use case I see for supporting pickle is to allow parallelizing task execution with the multiprocessing module.

Relevant SO question: Pickling dynamically generated classes?