This helps minimize the fragmentation that occurs when 2MB imports are
interleaved across NUMA domains. This complements D24548, which aims to
reduce internal fragmentation caused by vmem's unwillingness to coalesce
across imported spans.

This is based upon the presumption that potential overallocation is not
harmful on NUMA systems, i.e., we are at worst wasting page table pages
used to extend the kernel pmap.