asyncpg __del__ methods call asyncio APIs directly, which isn't guaranteed to work and indeed, doesn't work

asyncpg has several `__del__` methods that call into asyncio APIs. But... you can't safely call into asyncio APIs from a `__del__` method, because `__del__` methods can be run in arbitrary threads, or at arbitrary moments when the loop's internal data structures are in an inconsistent state.

Discovered here: https://github.com/python-trio/trio-asyncio/issues/44

Looking at the traceback in that issue, I suspect that the reason this hasn't been noticed before is that the `__del__` method ultimately ends up calling `loop.call_soon(...)`. And if you do that from another thread, then with regular asyncio, things will mostly seem to work – it won't actually wake up the loop the way `call_soon_threadsafe` would, but if the loop is still running then it will eventually get run, and the default `call_soon` will not explode or otherwise notice if it's called from the wrong thread. But in that issue, someone's using asyncpg with an alternative asyncio loop that fails when `call_soon` is called from the wrong thread, and this breaks asyncpg.

I guess asyncpg should go through all its `__del__` methods and wrap them in a `loop.call_soon_threadsafe`?

Start by inventorying asyncpg's __del__ methods and the loop.call_soon paths implicated by the linked Trio issue. Check each cleanup path against the thread-safety and loop-state concerns described there. Done means the affected cleanup behavior works with alternative asyncio loop implementations without unsafe direct calls.

Written by the indexing model from the issue text.