amd64/pmap: fix user page table page accounting

When a superpage mapping is destroyed and the original page table page
containing 4KB mappings that was being held in reserve is deallocated,
the recently introduced user page table page count was not being
decremented. Consequentially, the count was wrong and would grow over
time. For example, after multiple iterations of "buildworld", I was
seeing implausible counts, like the following:

vm.pmap.kernel_pt_page_count: 2184
vm.pmap.user_pt_page_count: 2280849
vm.pmap.pv_page_count: 106

With this change, I now see:

vm.pmap.kernel_pt_page_count: 2183
vm.pmap.user_pt_page_count: 344
vm.pmap.pv_page_count: 105

Reviewed by: kib, markj
Differential Revision: https://reviews.freebsd.org/D33276