Include Dask Support

[jakirkham]

Would be great if Dask Array's could be an optionally supported input/output type. This way one could construct a graph of the computation using rank_filter with proper haloing and compute in parallel when ready.

[jakirkham]

Alternatively it might be reasonable to just use multiprocessing.dummy.Pool, which would give us a way to run the filter in multiple threads and avoid any additional dependencies.

Edit: Can probably do something like this.

multiprocessing.dummy.Pool().map_async(<func>, <iterable>).wait()

Edit 2: This wound up being slower than not using a thread pool. The overhead of scheduling and lock contention must be making this not viable.

[jakirkham]

As a step in this direction, the filter now releases the GIL for the entirety of the for-loop.