Avoid sync deadlocks when the default IO thread has stopped - #2210

The intent is to make sync() a no-op during shutdown here; so what happens at the call site - a normal exception?

Good catch - the quiet return was wrong. A caller would get None and treat it as a real result: one wrapper hands None back to its caller, another would loop forever, and the HTTP code would crash on it later. The only place that needs quiet at shutdown is close_session. So now sync() always raises a clear RuntimeError when the IO thread is gone, close_session catches it and shuts down quietly, and the tests expect the raise in both cases.

HTTP may be where we exhaustively test AsyncFileSystem, but there are other implementations in other repos (s3fs, gcsfs. adlfs and maybe other). Would they need a change?

Good question - I only changed the HTTP one here, so I checked the others by reading their main branch. s3fs and gcsfs call sync() from close_session only when the loop reports running, but they catch just FSTimeoutError - so if the IO thread is gone but the loop still looks alive, they'd now see this RuntimeError instead of falling back to closing the connector. Worth them catching RuntimeError the same way. adlfs calls sync() directly from its weakref finalizers with no try/except, so at shutdown it would now raise from the finalizer instead of silently returning - a matching guard is a one-line follow-up in adlfs, not this PR. I haven't run any of them against this branch, so that's a reading of the code, not a test result.

How about (loop._thread_id or 0) == iothread[0].ident? That would account for any supplied loop.

Good direction - a supplied loop that lives on the fsspec IO thread should get the same guard. Applied in 9f79f44 with two adjustments, both verified locally:

1. Kept the identity check as a disjunct. After a clean loop.stop(), run_forever's finally clears loop._thread_id back to None, so a thread-id-only check misses exactly the stopped-default-loop case this PR is for (verified here: _thread_id equals the thread's ident while the loop runs, None after stop).
2. getattr(loop, "_thread_id", 0) because uvloop loops don't carry _thread_id at all, so the attribute access would raise for them.

The check is now:

if iothread[0] is not None and (
    loop is globals()["loop"][0]
    or (getattr(loop, "_thread_id", 0) or 0) == iothread[0].ident
):

Tests: added a case for a supplied loop whose _thread_id matches the stopped IO thread's ident (raises), and gave the custom-loop mock an ident. The dead-thread tests pass; the rest of test_async.py shows the same 5 environment failures as unpatched master (no aiohttp on this box), nothing new. Ruff format clean.