[minor] Test helper fill_socket is probably unreliable on Windows

Suppose you want to set up a socket where send blocks. It's helpful for testing; we do it in several places. Hence this helper from trio/_core/tests/test_io.py:

def fill_socket(sock):
    try:
        while True:
            sock.send(b"x" * 65536)
    except BlockingIOError:
        pass

Except... as documented by @vstinner in this epic bit of debugging, apparently on Windows this can occasionally fail. (The send raises BlockingIOError... but if you call send again afterwards, then it succeeds, even if no-one has called recv.)

This hasn't bitten us yet, but I guess it probably will eventually. (Or maybe it has, but our tests are written in such a way that they accidentally passed?)

Here's Victor's current solution: https://github.com/python/cpython/pull/8327/files

Start in trio/_core/tests/test_io.py at fill_socket and inspect its usages in the tests. Read the linked CPython issue and pull request for the Windows behavior, then run the affected tests on Windows if available. Done means the helper reliably creates a socket whose send blocks without allowing a later send to succeed unexpectedly.