Unbreak the nmi handler (again): I placed INTR_CLEAR_GPRS in the wrong

Unbreak the nmi handler (again): I placed INTR_CLEAR_GPRS in the wrong
place and GS.base was horked on return.  Also, the frame passed to ddb
didn't have the %rbp<-->tf_err swap, which would have confused backtraces.
Now if we can just come up with a way to automate testing the NMI handler
with qemu...