... or so I hope. The main observation was that we still have both
hash_vcpu_for_each() and hash_domain_for_each(), where the latter was
introduced in 2014/15 to replace the former. Only some eight years
later we can now complete this conversion. Everything else addresses
other things noticed along the road.

Note that patch 8 conflicts contextually but not functionally with
patches 3 and 4 from "x86/mm: address aspects noticed during XSA-410
work".

1: replace sh_reset_l3_up_pointers()
2: drop hash_vcpu_foreach()
3: rename hash_domain_foreach()
4: drop a few uses of mfn_valid()
5: L2H shadow type is PV32-only
6: x86/shadow: re-work 4-level SHADOW_FOREACH_L2E()
7: reduce effort of hash calculation
8: call sh_detach_old_tables() directly
9: harden shadow_size()

Jan