While the second phase of select is running we queue send/recv cases to
corresponding channels. At some point - when some of the cases are
already queued - a peer goroutine might try to send/recv on that
channel. And it will succeed because a waiter was queued to the channel.

At the same time select is continuing its enqueue loop and before enqueuing
to a channel it tries to send/recv there. If that channel became just ready
(i.e. just after select poll phase) the try to send/recv will succeed. This
means that actually 2 select cases could be executed at the same time.

Fix it by carefully checking whether some case already won before trying
to send/recv on a channel.

This fixes the test failures that were demonstrated by previous 2 patches.