HandleTable mutated without the GIL via _pm_release (suspected race) #117

Found while investigating #98 (by reading, not observed): UpcallTrampoline.releaseObject, reached through ctypes _pm_release (which releases the GIL around the foreign call), mutates HandleTable without the GIL, while AsyncUpcall.kt states every HandleTable mutation happens under the GIL. Under free-threaded or concurrent use this can corrupt the table.

Completion: either prove the call holds a lock that makes it safe (and document it), or take the GIL / the table's own lock on that path; a concurrency stress test (many threads releasing and registering handles) passes under -PpythonFreeThreaded=true and default.