tests: 5s wall-clock time_limit flakes under xdist (test_ipv4_unit truncated option) #1058

tests/protocols/internet/test_ipv4_unit.py::IPv4UnitTests::test_a_truncated_option_still_parses_its_declared_length failed on #1054's Python 3.12 leg (job 112038796404) with TimeoutError: did not finish within 5s from tests/_support.py:72. The other four Python legs passed on the same head, and #1054 touches only httpv1.

The traceback ends inside nested <frozen abc> __subclasscheck__ calls, reached from infoclass.__update__'s isinstance(dict_, collections.abc.Mapping) (infoclass.py:455). So the parse was slow, not hung. That fits ABC subclass caches growing as purge_modules re-imports pcapkit within one xdist worker; #1052's investigation found the same re-import growth on the unittest leg. The mechanism is UNVERIFIED.

time_limit(5) is wall-clock (signal.alarm) and is used 22 times under tests/, so any of them can flake on a loaded 4-core runner.

Fix: keep the hang guard, but make it robust to a slow runner. Raise the per-test deadline, or derive it from a measured baseline, and confirm whether ABC cache growth from repeated re-imports is the real cost. This overlaps with #1052's item B (run_unittest_leg re-import bounding) and item D (pytest-timeout).