…rpc#40447)

This PR changes the logic of `shutdown_grpc_aio` to skip `_actual_aio_shutdown` python interpreter is already [being finalized](https://docs.python.org/3.14/glossary.html#term-interpreter-shutdown) (cleaning up resources, destroying objects, preparing for program exit, etc). `_actual_aio_shutdown` involves `PollerCompletionQueue` shutdown, followed by core [`grpc_shutdown`](https://grpc.github.io/grpc/core/grpc_8h.html#a35f55253e80714c17f4f3a0657e06f1b) API call.

Reasoning:
1. During finalizations, in come cases resources we're accessing may already be freed, and the order is not deterministic. Some of the resources being unloaded prior the  `_actual_aio_shutdown` call: `_global_aio_state`, `AsyncIOEngine` enum, or even python libraries like `sys`. This leads to errors like `AttributeError: 'NoneType' object has no attribute 'POLLER'`.
2. `PollerCompletionQueue.shutdown()` will try to wait on its poller thread to finish gracefully. In py3.14, `PythonFinalizationError` is raised when `Thread.join()` is called during finalization. I think the logic here is similar to (1): these threads may have already been deallocated.

Note that in some cases users were able to prevent `_actual_aio_shutdown` from being called by manually calling `init_grpc_aio` prior to initializing any grpc objects.  This resulted in an incorrect positive refcount, which prevents `_actual_aio_shutdown` from being run. Before the above finalization check was added this side-effect was sometimes misused to avoid deadlock on finialization (grpc#22365).

This PR:
- Fixes grpc#39520
- Fixes grpc#22365
- Fixes grpc#38679
- Fixes grpc#33342
- Fixes grpc#36655

Closes grpc#40447

COPYBARA_INTEGRATE_REVIEW=grpc#40447 from sergiitk:fix/aio/shutdown 11114f6
PiperOrigin-RevId: 804971756