Change the way unmounting happens to guarantee that the
client programs are allowed to finish up (coda_call is
forced to complete) and release their locks. Thus there
is a reasonable chance that the vflush implicit in the
unmount will not get hung on held locks.