test(support): time_limit's one-shot alarm can be swallowed and hang an xdist worker #1892

tests/_support.time_limit (tests/_support.py:75) arms a single signal.alarm(seconds), and its handler raises TimeoutError exactly once. #1891 found that this one-shot pattern can be lost. A finaliser that runs the raise sends it to sys.unraisablehook, and an except OSError swallows it because TimeoutError is an OSError (Extractor.__del__ catches exactly that). Either way the body then runs until pytest-timeout's 900s backstop ends the xdist worker. That is what hung the #1428 deadline test (#1889).

time_limit has the same exposure: a lost raise turns a bounded test into a 15-minute worker crash with no traceback.

Fix: once it has fired, re-arm until the with block exits, as #1891 does for the generator's _deadline. The re-arm must not outlive the block, and the previous handler and pending alarm must still be restored. Add a test where a finaliser swallows the first raise, and check that it fails quickly without the fix.

Related: under xdist, pytest-timeout writes its stack dump to the worker's stdout, which is /dev/null. A hang like #1889's leaves only node down: Not properly terminated. Consider sending the dump somewhere the job log keeps, for example through faulthandler to stderr or a file the job uploads.