Almost fixes #8753 (which still needs a test).

Otherwise, a null closure could be executed to reclaim, the reclaimation
finished edge would never be executed, and a future allocation will
stall forever.

1. In the State change test, 123->0 will hit a c-core GPR_ASSERT if
debugging is enabled because 123 doesn't map to a state that can
be printed.
2. In the Server test, 8888 is sometimes unavailable, so we just use
a wildcard port.

The server cleanup method is untested.

The join() function that exposes it is only called by the internals of threading.py, and we don't hold a reference to the server thread to explicitly join() it, and I'm not sure we should add a reference just for this purpose.

Moreover, the threading.py only calls join(None), the code path in question isn't even exercised by the internals of threading.py. Its just there to make sure we properly follow the join(timeout) semantics.

Previously, we'd keep retrying a connection until the channel closed.
With this change, we only retry connecting *if there's a watcher on the
subchannel connection state*. This ultimately means that if the
lb_policy doesn't care if the subchannel connects, it'll stop trying.

- we were treating an int as atomic, which is dubious at best
- it was possible to shutdown while an accept was being handled, and
process that shutdown accept before the real accept finished, leading to
a use-after-free up the stack

It can happen that a timer event causes something to be queued to an
IOCP, which means that on Windows we need to flush that queue each time
a timer event fires during shutdown.

Localize global_wakeup_fd declarations, instead of trying to share them,
so that this bug is less likely to occur in the future.

This occurred when running a QUIC test in an IPv6-only environment:

E1123 16:04:20.282041  159850 heap-profile-table.cc:489] RAW: Leak of 1360 bytes in 17 objects allocated from:
        @ 0x7f160b8923ab gpr_malloc
        @ 0x7f160b8cff27 copy_error_and_unref
        @ 0x7f160b8d0151 grpc_error_set_str
        @ 0x7f160b8e3ff8 error_for_fd
        @ 0x7f160b8e3edd grpc_create_dualstack_socket
        @ 0x7f160b8ea5c3 grpc_udp_server_add_port
        @ 0x7f1627a65d20 grpc_server_add_quic_port
        @ 0x7f1627a6bea4 grpc::(anonymous namespace)::QuicServerCredentialsImpl::AddPortToServer()
        ...

Add use_tls, use_test_ca, and server_host_override flags to Python
stress client. These are needed to run the stress client against a local
server that is using tls.