Deserialization does not solve deprojection a second time; a deprojected RoiSet is serialized as 3D.  Tight patch mask format depends on dimensions of ROIs themselves, so 2d PNGs are possible in a 3D set if no ROI is thicker than 1

Applied same optimization to binary mask computation, but this still his memory issue on large dataset

Gradually handling three different cases of RoiStack: pure 2d, pure 3d, mixed (3d stack, 2d deprojected object IDs)

Parameterized option to create unique subdirectory for each RoiSet, but this still isn't propagating to dataframe called in .serialize

# Conflicts:
#	model_server/base/api.py
#	model_server/base/roiset.py
#	tests/base/test_roiset_pipeline.py

RoiSet.run_exports and .get_export_product_accessors no longer take a channel directly, but rather pull these from parameters of individual patch products

Requirements for building on cluster

See merge request rhodes/model_server!79

# Conflicts:
#	tests/base/test_roiset_pipeline.py

Updated tests: smoothing is now more aggressive, lowered threshold of IOU for comparison; task endpoints updated

Updated tests: smoothing is now more aggressive, lowered threshold of IOU for comparison; task endpoints updated