kernel: queue: k_queue_poll: Fix slist access race condition

All sys_slist_*() functions aren't threadsafe and calls to them
must be protected with irq_lock. This is usually done in a wider
caller context, but k_queue_poll() is called with irq_lock already
relinquished, and is thus subject to hard to detect and explain
race conditions, as e.g. was tracked in #4022.

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