Add pika::wait

[msimberg]

Fixes #696.

Calls thread_manager::wait. pika::wait can also be called on a pika thread, in which case it'll ignore the calling thread and wait for all other work to finished. The behaviour is undefined/will likely deadlock if pika::wait is called concurrently from multiple pika threads.

Note, I made wait return an int to be consistent with the other related functions, but I opened #703 to possibly change it to void.

[msimberg]

bors try

[bors]

try

Build failed:

• ci/circleci: inshpect

[msimberg]

bors try

[bors]

try

Build failed:

• jenkins/cscs-daint/clang-12-debug

[msimberg]

This seems to have exposed some problems in thread_manager::wait. I will try to get back to this in a few weeks.

[msimberg]

This is on hold because I had forgotten about the probabilistic nature of checking if the runtime is idle. We don't have a global indicator to check that all of the thread pools and MPI/CUDA polling is idle, which means that it's inherently racy to individually check those for idleness and infer idleness of the runtime from those.

We may need to try to introduce a global counter but we have to be careful not to kill performance with a global atomic counter.

[msimberg]

I've removed the test from this PR and left only the pika::wait wrapper for thread_manager::wait so that we can start replacing uses of thread_manager::wait in DLA-Future with a public API function. The test is in #704 but still needs a deterministic check for the runtime being empty.

[aurianer]

LGTM thanks!

[msimberg]

bors merge

[bors]

Build failed:

• jenkins/cscs-daint/clang-12-release

[msimberg]

Merging despite failures, which seem to mostly be timeouts on header tests.