fix: join a thread whose shutdown was only requested

[nicolas-grekas]

state.ShuttingDown means two things today and the code treats them as one. Normally phpThread.shutdown() is what publishes it, and it owns the rest of the sequence: close drainChan, wait for Done, force-kill past the grace period. But a worker that fails to boot publishes it from inside the PHP thread, at a point where that thread still has its whole C exit path ahead of it, down to go_frankenphp_on_thread_shutdown() which is what sets Done.

That publication also unblocks initWorkers(), so Init() fails and calls shutdown(), where RequestSafeStateChange() refuses the thread that is already shutting down and returns right away. drainPHPThreads() then releases php_main() into sapi_shutdown() and tsrm_shutdown() with a live PHP thread still registered in TSRM, which is exactly what the comment above that teardown says cannot happen, and it nils phpThreads while that thread still indexes it from its callbacks.

shutdown() now tells a shutdown that was requested from a thread that has exited: the grace period and force-kill wait moved into phpThread.waitForExit(), the owning path waits for Done as before, and the refused path waits for Done or Reserved on a single observation, so a thread that never booted returns at once and one that is merely shutting down is joined like any other. No path changes except a thread genuinely mid-shutdown at drain time, which today only the boot-failure path produces.

TestInitJoinsAThreadStuckInStartupTeardown holds a failing worker inside its shutdown hook and asserts Init() cannot return while the thread is held. Without the fix it reports the thread still shutting down, and the race detector flags drainPHPThreads() nil'ing phpThreads under the live callback.

Reported by @henderkes in #2617, where the background worker boot-failure path inherited it: #2617 (comment)