Fix all the things by fixing multiprocessing

[stephenbach]

Mutliprocessing in the helicopter view is the source of a lot of inconsistent bugs across OSs and Python versions. This PR fixes promptsource when using the spawn mode (default in python 3.8+ and required to run smoothly on latest OS X). It also changes it to the use spawn always so that hopefully there is more consistent behavior.

Before merging, we should test on a range of OSs and Python versions.

The main changes are actually pretty small, but the diff is huge because most of app.py had to be wrapped in a function. (This is so that this code is not executed in each subprocess, which was our problem when spawning.) The main changes besides wrapping the code is

• pass all_infos to get_infos as an argument
• Force start mode to spawn

[stephenbach]

Fixes #572
Fixes #588
Fixes #416
Fixes #584

[tianjianjiang]

A shameless self-promoter here: #411. It is almost like this PR without the UI changes. I've tested with python 3.8 and 3.9 on both macOS 10.15 Catalina and WSL.

[VictorSanh]

@stephenbach LGTM!

I have tested on MacOs Catalina with Python 3.7, 3.8 and 3.9, and everything looks good.
I also tested on Unix with python 3.7 for sanity check, same.

NB: I had to change L184 to split_sizes = {k: v.num_examples for k, v in subset_infos.splits.items() if hasattr(v, "num_examples")} because downloaded infos seemed to have change for the zaid/coqa_expanded and zaid/quac_expanded. We can also remove L186-L188's comments since they don't apply anymore.

Maybe keep the if infos" condition as sanity

[tianjianjiang]

I've tested using "datasets==1.18.4" without #735.
(With #735, "datasets==2.0.0" works, see https://github.com/tianjianjiang/bigscience-promptsource/runs/5802249762?check_suite_focus=true#step:4:26)

It works with "python>=3.7,<3.10" for macOS BigSur 11.6.5 on Intel, too.
I will also test Monterey and re-test Catalina soon, but they probably won't cause any issue.

For "python>=3.10,<4", see #733 (comment).