Kick off the suspend sequence from the keyboard in a SWI rather than
in the hardware interrupt context (even if it is likely just an
ithread). We don't document that suspend/resume routines are run from
such a context and some of the things that happen in those routines
aren't interrupt safe. Since there's no real need to run from that
context, this restores assumptions that suspend routines have made.

This fixes Thierry Herbelot's 'Trying to sleep while sleeping is
prohibited' problem.