trio.sleep_forever should mention that it raises if rescheduled

[A5rocks]

This causes a regression for anyone who was using trio.sleep_forever combined with trio.lowlevel.reschedule. Okay, maybe they (we (I)) should have been using trio.lowlevel.wait_task_rescheduled but the RuntimeError should have been phased in with a warning over one release cycle.

Originally posted by @richardsheridan in #3096 (comment)

[A5rocks]

I think this regression is ultimately ok, especially since it's already done. We can update the docstring to mention this as in #3096 (comment).