fix(as_llm): close owned OpenAI clients safely - #725

[P2] Prevent lazy model construction throughout shutdown

_release_model() clears _model in its separate task before the outer close() resumes and sets _closed here. A concurrent llm.model access in that window constructs a second owned client. Shutdown then returns with is_started=False, _closed=True, and _owns_model=True, and the second client's close() was never awaited. The getter also keeps returning that model after shutdown because it checks _model before _closed.

I reproduced this by setting an asyncio.Event in the first client's async close method and having another task wait for that event before reading llm.model. Please make the transition into shutdown prevent new model construction before provider cleanup can yield, coordinate that transition with the lifecycle lock, and add a regression test for this interleaving.

[P2] Do not clean up another lifecycle operation's model when a startup waiter is cancelled

super().start() can raise CancelledError while this invocation is still waiting to acquire _lock. This handler then calls _close() without owning that lock, even though another startup or shutdown operation owns the model. With a running component whose client close is blocked, queueing a start() and cancelling it causes client.close() to run twice concurrently. With one startup blocked and a pre-created model, cancelling a second queued startup closes and discards the first startup's model while that startup is still running.

Please distinguish cancellation before acquiring the lifecycle lock from failure within this startup invocation, and serialize rollback with the other lifecycle operations. A cancelled lock waiter should not tear down their resources.