Fixed deadlock in injection_queue_depth_multi_thread test

[jofas]

This closes #6847 by removing the deadlock from the test as discussed in #6876. The changes I made are based on this suggestion. I made the blocking tasks block by letting them wait on a message from a std::mpsc channel. This is quite convenient IMO, because we can exit the blocking tasks simply by dropping all Senders of the channel.¹

Footnotes

1. Unlike the previous implementation that used Barriers for synchronization—even if the test were to unforeseeably panic—the blocking tasks would still exit and cause the test to fail early instead of timing out. ↩

[Darksonn]

We usually do this kind of retrying by defining two functions, to separate out the concerns.