Support pickling sources.

[manthey]

This also supports pickling tiles from the tile iterator.

Closes #1069.

[banesullivan]

I'm curious what's driving the need for this. Multiprocessing?

[manthey]

I'm curious what's driving the need for this. Multiprocessing?

Yes -- @annehaley tried to do some multiprocessing and the inability to pickle sources made the book keeping more involved. This would also make the multiprocessing cleaner in the HistomicsTK examples.

[manthey]

For example, this:

   for tile in ts.tileIterator(**it_kwargs):
        tile_position = tile['tile_position']['position']
        cur_nuclei_list = dask.delayed(detect_tile_nuclei)(
            args.inputImageFile,
            inputImageArgs,
            tile_position,
            it_kwargs,
            args, src_mu_lab, src_sigma_lab
        )
        tile_nuclei_list.append(cur_nuclei_list)

becomes:

   for tile in ts.tileIterator(**it_kwargs):
        cur_nuclei_list = dask.delayed(detect_tile_nuclei)(
            tile,
            args, src_mu_lab, src_sigma_lab
        )
        tile_nuclei_list.append(cur_nuclei_list)

where, since tile is pickleable (since the source is pickleable), it can be passed by itself rather than as a file name, tile position, and arguments for the tile source and tile iterator.

And the delayed function, instead of doing

ts = large_image.getTileSource(slide_path, **slide_args)
tile = ts.getSingleTile(
        tile_position=tile_position,
        format=large_image.tilesource.TILE_FORMAT_NUMPY,
        **it_kwargs)

just uses a reference to tile

[banesullivan]

Thanks for clarifying -- good to capture these examples too