In vmap::pagefault the assert assumed a PRIVATE vma would have a
vmnode with 1 ref.  In vmap::copy, however, the vmnode reference is
incremented, then the vma is replaced with one of type COW.  If a
thread is pagefaulting while another is forking there is a race.

Previously, sched always paused call tracing but if it then returned
directly because there was nothing else to schedule, call tracing
didn't get resumed.  We could simply resume tracing in this case, but
this fix instead moves the pause *after* the next proc is picked, so
that we don't get lots of spurious pause/resume pairs.

Fix some compile errors (mtstart depended on C's implicit cast from
function pointers to void* and user.h lacked some extern "C" function
declarations mtrace needed) and some logic errors (mtstart wasn't
happening in idle procs and there was a perplexing mtpause going out
of sched).

Now they agree between kernel, user, and Linux.  (This was motivated
by link errors in mtrace caused by disagreements between xv6 kernel
and user code, but also seems like the right thing to do.)

We panic all over the place on OOM, but it'd be good to slowly work
towards slightly saner error handling.

Somewhat cleaner than hitting a page fault.

..now xdu and xls work in xv6 userspace with the wq stuff..