Fix issue #2135

[prestonhale]

Fully fixes the issue in #2135 for me. Not sure what the purpose of maintaining the user_dispatcher was in the first place so take this with a grain of salt. If you do want to dig deeper the issue seems to be the _dispatch generator.

[cyberw]

It seems to break a test, but maybe the test is wrong?

[mboutet]

@cyberw see #2152 for the appropriate fix and reproducing test.

As for test_stop_timeout_with_ramp_down, it is an odd test indeed. Even when it was introduced in ded61e0, the test name doesn't make much sense. It does not strike me as test for validating the stop_timeout as the name implies. If it was the case, I would expect the task to sleep for more than two seconds.

Moreover, the docstring I added in d9bb87f is outdated and misleading. I'll remove it in a separate PR.

However, I'm not sure it should be removed as the test itself makes sense (apart from its name).

[cyberw]

This should be closed then, I suppose.