kernel: k_poll: Introduce separate status for cancelled events

Previously (as introduced in 48fadfe6), if k_poll() waited on a
queue (or subclass like fifo), and wait was cancelled on queue's
side using k_queue_cancel_wait(), k_poll returned -EINTR. But it
did not set event->state field (to anything else but
K_POLL_STATE_NOT_READY), so in case of waiting on multiple queues,
it was not possible to differentiate which of them was cancelled.

This in particular broke detection of network socket EOF conditions
in POSIX poll() implementation.

This situation is now resolved with introduction of explicit
K_POLL_STATE_CANCELLED state, which is now set for cancelled queue
(-EINTR return remains the same).

This change also elaborates docstring for the functions mentioned, to
document this behavior.

Fixes: #9032

Signed-off-by: Paul Sokolovsky <paul.sokolovsky@linaro.org>