Add init_worker_code for code that runs once per worker

[christiangnrd]

The GPUArrays testsuite takes at least 3s to load. It saves a lot of time to only load it once per worker instead of once per test especially in the memory constrained runners that are 7-8GB Mac Minis

Non-breaking

[vchuravy]

We did intentionally removed this functionality at some point during a refactor.

Would it make sense to move the GPUArraysTest instead into a package?

[christiangnrd]

Would it make sense to move the GPUArraysTest instead into a package?

With the idea being that a separate package would be precompiled so loading the precompiled test suite for every test would introduce less overhead?

I think that would add a significant amount of friction for contributing to GPUArrays.

Maybe I de-emphasize it in the docs and make clear that init_code is much simpler and should be the default outside of large initialization code?

[christiangnrd]

@giordano I've addressed your feedback. It should be more clear that this is only really meant for situations with long-running initialization code, and I fixed the test that initially didn't test anything

[vchuravy]

I think that would add a significant amount of friction for contributing to GPUArrays.

It would not need to be a seperate repository, and I don't even think we would need to register it.

[maleadt]

LGTM. One nit above.

[christiangnrd]

Should I bump to 2.2 here or is it better as a separate commit/PR?

[maleadt]

You can bump here

[christiangnrd]

Fixed up and added a note in the documentation