On a machine with lots of CPUs, it's is very possible to generate port collision by simply running to many tests. The port computation will wrap up and collide with a test which is still running. To avoid this, we change to strategy to use port skips by range depending on the CPU count.
While at it, I also removed the global and renamed it for clarity, wrapping it into a class to make mypy happy (function attributes are not supported but class attributes are).

On a 32 core machine:

./test/functional/test_runner.py