Tests trying to find unused ports are flakey when automated

• uvloop version: 0.17.0
• Python version: 3.10
• Platform: nix/darwin
• Can you reproduce the bug with PYTHONASYNCIODEBUG in env?: N/A
• Does uvloop behave differently from vanilla asyncio? How?: N/A

The tests using find_free_port (like https://github.com/MagicStack/uvloop/blob/d2deffefa18653636eb03ea77a5dab6e4febf6c6/tests/test_tcp.py#L252) may fail in busy CI environments. For example building for nixpkgs involves running tests every time. It ends up with: (https://hydra.nixos.org/build/197525892/nixlog/1)

=================================== FAILURES ===================================
_______________________ Test_UV_TCP.test_create_server_5 _______________________
Traceback (most recent call last):
  File "/private/tmp/tmp.xkcj6kx4o4/tests/test_tcp.py", line 245, in test_create_server_5
    self.loop.run_until_complete(runner())
  File "uvloop/loop.pyx", line 1517, in uvloop.loop.Loop.run_until_complete
  File "/private/tmp/tmp.xkcj6kx4o4/tests/test_tcp.py", line 238, in runner
    srv = await self.loop.create_server(
  File "uvloop/loop.pyx", line 1790, in create_server
OSError: [Errno 48] error while attempting to bind on address ('::', 50000, 0, 0): address already in use
_______________________ Test_UV_TCP.test_create_server_6 _______________________
Traceback (most recent call last):
  File "/private/tmp/tmp.xkcj6kx4o4/tests/test_tcp.py", line 271, in test_create_server_6
    self.loop.run_until_complete(runner())
  File "uvloop/loop.pyx", line 1517, in uvloop.loop.Loop.run_until_complete
  File "/private/tmp/tmp.xkcj6kx4o4/tests/test_tcp.py", line 255, in runner
    srv1 = await self.loop.create_server(
  File "uvloop/loop.pyx", line 1790, in create_server
OSError: [Errno 48] error while attempting to bind on address ('::', 50000, 0, 0): address already in use
=============================== warnings summary ===============================

Do those tests need to use a common ipv4/6 socket? Could they use only one protocol at a time with port 0 and then get the local port number from the result?

Or maybe build the retry loop around the first socket creation rather than 2 independent steps?

Start in tests/test_tcp.py, especially the tests using find_free_port around test_create_server_5 and test_create_server_6. Reproduce the failures in an automated or busy environment, then evaluate using port 0, a shared IPv4/IPv6 socket, or retrying the first socket creation. Done means the tests no longer fail because the selected port is already in use.