Fix transient timeout failures

[danielsf]

Over the last two days, I've seen the transient timeout failures resurface in ophys_etl_pipelines. See for instance:

https://app.circleci.com/pipelines/github/AllenInstitute/ophys_etl_pipelines/2417/workflows/f065ff74-74f7-45dd-ba5e-2cb4d807c14b/jobs/7870

I think that the failures always arise from code that invokes some variation on multiprocessing.Pool.

I cannot say for certain why CircleCI has a problem with multiprocessing.Pool. I would like to propose we solve the problem by replacing Pool with by-hand invocations of multiprocessing (i.e. directly starting Process instances, and calling join on them as necessary). I do not believe code that uses multiprocessing that way has ever provoked this failure.

Modules that use multiprocessing.Pool include

sine_dewarp
segmentation.modules.calculate_edges
dff
qc.video.correlation_graph

These are all of the modules that were giving me trouble when I was transitioning our CircleCI builds over to Docker containers that we built.

I am happy to discuss this further with anyone who thinks I am mistaken.

[danielsf]

I believe this is another example of the same failure

https://circleci.com/api/v1.1/project/github/AllenInstitute/ophys_etl_pipelines/7949/output/103/0?file=true&allocation-id=62103bc229707648f308e497-0-build%2F627D31A2