🐛 fix(run): break deadlock in execution interrupt chain

[gaborbernat]

On Windows CI (~1/40 runs), a subprocess can hang indefinitely during environment setup — either in virtualenv's interpreter discovery (Pattern A) or during package installation/provisioning (Pattern B). Analysis of the last 30 days of CI revealed 18 flaky timeout failures across 9 different tests, 89% on windows-2025 and 11% on macos-15. 🪟 The affected tests are not specific — any test that runs tox in-process where a subprocess hangs triggers the same deadlock.

The root cause is an unbreakable deadlock chain in common.py. thread.join() blocks the main thread indefinitely so signals from pytest-timeout can never be delivered. as_completed() blocks the tox-interrupt thread so it can never check the interrupt event. And executor.shutdown(wait=True) prevents done.set() from firing even after an interrupt is acknowledged. For Pattern B, tox_env.interrupt() would kill the hung subprocess since it's tracked in _execute_statuses, but it can never fire because KeyboardInterrupt can't reach the blocked main thread.

Fix	Pattern A (discovery)	Pattern B (install/provision)
thread.join(timeout=1) loop	Allows signal delivery	Allows signal delivery
_next_completed with interrupt check	Breaks out of wait	Breaks out of wait
executor.shutdown(wait=not interrupted)	Skips stuck worker	Skips stuck worker
daemon=True on thread	Process can exit	Process can exit
done.wait(timeout=5)	Bounded cleanup	Bounded cleanup

⏱️ The blocking as_completed() is replaced with a polling _next_completed() that checks the interrupt event every second via concurrent.futures.wait(timeout=1, return_when=FIRST_COMPLETED). The interrupt thread is made daemon so the process can exit if it's stuck. thread.join() uses a timeout loop so signals can be delivered on Windows (where lock.acquire() without timeout ignores signals). The interrupt handler gets bounded waits so cleanup doesn't hang forever.

For Pattern A, the upstream fix is in tox-dev/python-discovery#42 which adds a 5s timeout to process.communicate() in _run_subprocess. Together, these changes eliminate both hang patterns.

[rahuldevikar]

We should do similar change to LocalSubprocessExecuteStatus.wait() in __init__.py so it periodically yeilds control and checks for interrupts instead of blocking indefinitely on WaitForSingleObject