Close path: concurrent close_async finalizes twice, and three untested guarantees #616

This was generated by AI during a third pre-4.0 audit of main at f6e343d.

One bug and three untested guarantees in the close path. Mutation testing changed each of items 2 to 4 in the code and all 989 tests still passed. TDD for each.

1. Concurrent close_async() finalizes twice. Repro:

   async def fin(r: R) -> None:
       log.append("start"); await asyncio.sleep(0.01); log.append("end")
   class G(Group):
       r = Factory(R, cache=CacheSettings(finalizer=fin))
   c = Container(groups=[G]); c.resolve(R)
   await asyncio.gather(c.close_async(), c.close_async())
   # log == ['start', 'start', 'end', 'end']

   Fix: detach the creation order before the first await (or mark items finalized before awaiting), and state in design-decisions whether concurrent close is supported.
2. Finalizer ownership. resolver_compiler.py:120 appends to target._creation_order. Changing it to container._creation_order passes the suite, yet then an APP singleton first resolved from a REQUEST child is finalized when the child closes and rebuilt afterwards, contradicting docs/providers/lifecycle.md:191. Test: resolve an APP item with a finalizer through a child, close the child, assert the finalizer did not run and the singleton is the same object.
3. Order of leftover async finalizers. Removing remaining.reverse() in cache.py close_sync passes the suite. Test: two async finalizers, sync close, then async close, assert newest first.
4. No double finalize on the async path. Removing cache_item.finalized = True in cache.py close_async passes the suite; test_cached_factory.py:69 covers only sync. Test the clear_cache=False case: close_async, reopen, close_async again.
5. Small: test_cache.py:114 races threads without the thread_race marker, so the free-threaded stress run skips it. Make just test-race select -m thread_race over the whole suite instead of listing files.