optimize dask graph computation

[Joachim Ungar]

collect multiple tasks into larger tasks in order not to overwhelm the scheduler:

• tile batches, no overviews (concurrent.futures, dask)
  no dependencies
• tile batches, with overviews (dask)
  dependencies between tile batches; use tile outputs within one tile batch
• tile batches with preprocessing tasks (dask)
  dependencies between all tasks; use preprocessing output for follow up (tile) tasks