Improve the locking "stuff"...

For regular vstatlock things, only hold the lock long enough to copy the
variables that will be used.

For vstat.vmem stuff, create a separate reference-counting vmem structure
and a new vmem_lock so the memcpy() of the vmem won't block basic vstat
accesses.

This appears to fix the recently-added "blockiness" of screen updates in
SyncTERM, without risking deadlocks or unlocked accesses... the major issue
holding up a release.