Report failures before the pytest hard-exit

[aleksUIX]

Summary

Any failing test run currently produces a bare F and exit code 1 with zero diagnostics, locally and in CI logs. The os._exit shutdown workaround in tests/conftest.py fires before pytest prints the FAILURES section, and it also drops buffered output. This PR keeps the workaround (the chromadb/posthog atexit hang it dodges is real) while restoring failure reporting.

Before / after

Before, on a clean checkout (pytest tests/unit/test_openrtb_parser.py -q):

........F....                                                            [100%]

That is the entire output. After:

........F....                                                            [100%]
=================================== FAILURES ===================================
____________________ test_round_trip_builder_then_parser_recovers_refs ____________________
...
E               RuntimeError: Could not locate ad_seller_system in path ancestry ...
=========================== short test summary info ============================
FAILED tests/unit/test_openrtb_parser.py::test_round_trip_builder_then_parser_recovers_refs
1 failed, 12 passed in 0.09s

Why two changes

1. Hook placement: the terminal reporter prints FAILURES, short summary, and the stats line after the non-wrapper pytest_sessionfinish hooks run, so hard-exiting there kills reporting. The exit now happens in pytest_unconfigure, which fires after all reporting; pytest_sessionfinish only records the exit status.
2. Flush before os._exit: os._exit skips stdio flushing, so with block-buffered output (file redirect, CI log capture) the report was lost even with correct hook order. Both streams are flushed first.

Verification

• Failing run: full traceback, short summary, and stats now print; exit code 1 (was already correct and stays correct, preserving the behavior from Fix CI workflow to fail Test job on pytest failures (ar-r82f.16) #15).
• Full unit suite: completes in ~6 seconds with no shutdown hang, so the Disable crewai shutdown telemetry hang (ar-r82f.21) #21 workaround is intact.
• ruff check clean.

The failing test shown above is a separate fresh-clone issue; a follow-up PR ports the skip pattern buyer merged in buyer-agent#92.