Remove the last source of SMP TLB invalidations in the critical code path

Remove the last source of SMP TLB invalidations in the critical code path
for fork/exec/exit/wait sequences.  The pmap page directory is now left
cached in kernel memory until the pmap is destroyed via the objcache
dtor function for VM spaces (vmspace_dtor() -> pmap_puninit()).

This improves fork/exec performance and significantly reduces the IPI
messaging load and other cross-cpu interactions.