I think the way forward is to write a uvmspace_purge function which
removes all (most?) userland mappings, and is called from an earlier
point in exit1(), when we can still sleep.  It leaves the map data
structure live and consistant, just without any mappings.

That results in the map being as empty as possible.  It probably copes
some of the top code inside uvm_map_teardown()

(20 years of worrying about pmemrange contention has resulted in a
different form of contention hurting instead).

The proposal is that the reaper sequence does not get changed at this
point.  However now the reaper will be faster because the map is has
fewer resources to return.

This is the first step, mappings removal in exiting-process context, and
NO CHANGES to the reaper or process sequencing.  

Once that works, the next step is probably to ensure the purge function
also cleans up the pmap resources correctly, in process context.

When that works, uvmspace_free/uvm_map_teardown probably is changed to
assert the user map / user pmap is actually empty.

Then the optimizations can begin.