Consume connections better in socket-level tests

[hodbn]

Add a utility that allows better timing of test responses, instead of socket consuming.

In Windows, after the second operation on a remotely closed socket a WSAECONNABORTED is raised.
In our socket-level tests, we implement a dummy HTTP server that only consumes data from the client socket, therefore it can't know when to stop consuming and might close the socket too early. This might cause unnecessary send/recvs on the already closed socket, which in turn will raise WSAECONNABORTED when it could be avoided.

I've added the 2 tests that are currently skipped on Windows (since #1951).
This change should prevent WSAECONNABORTED when testing the aforementioned tests on Windows.

[codecov]

Codecov Report

Merging #1958 into master will not change coverage.
The diff coverage is n/a.

@@            Coverage Diff            @@
##            master     #1958   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files           24        24           
  Lines         2187      2187           
=========================================
  Hits          2187      2187           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 0f11d38...edad878. Read the comment docs.

[pquentin]

@hodbn I merged master into your branch, since it was a prerequisite before merging this anyway

[pquentin]

This looks like black magic to me, but is great if it works! Would like to have another pair of eyes before I merge it

[hodbn]

This looks like black magic to me, but is great if it works! Would like to have another pair of eyes before I merge it

I've added another sentence of docs, hopefully the class purpose is more clear now.

[pquentin]

Let's try that! Thanks.

[PleasantMachine9]

I'm guessing this is the change that makes the CI flaky?
https://github.com/urllib3/urllib3/pull/1970/checks?check_run_id=1136185374

At least the MacOS tests failed 2/2 times, and first it was the 3.8, then the 3.6 test, and different ones, with the only common thing being that they were both socket-related (Connection Refused).

[hodbn]

I'm guessing this is the change that makes the CI flaky?
https://github.com/urllib3/urllib3/pull/1970/checks?check_run_id=1136185374

I don't think these failures are related to this change.
This change only affects test_preserve_chunked_on_redirect and test_preserve_chunked_on_broken_connection for now.