vmem: make SpaceReferenceMapping per-arch and fold space_aware_map dispatch

[danbugs]

Follow-up to review feedback on #1385. Bundling four related items into one issue because they form a single coherent change to the multi-space page-table walker interface and splitting them would produce awkward half-states.

Proposed changes

Move SpaceReferenceMapping into each arch module. Right now the struct lives in hyperlight_common::vmem, but it is only ever produced by arch::walk_va_spaces and consumed by arch::space_aware_map — the common module never touches it. Pulling it per-arch lets each architecture pick the representation that best fits its walker, and lets architectures that never emit AnotherSpace (amd64 today) use Void so the match arm in space_aware_map is statically unreachable via match impossible {} instead of today's unreachable! branch. Reviewer comment: #1385 (comment) and #1385 (comment).

Replace the depth: usize field with an arch-specific enum. On i686 the only valid depth is 1 (the leaf PT). On amd64 it would be one of { Pdpt, Pd, Pt }. A per-arch enum makes invalid depths unrepresentable and removes a class of off-by-one bugs. Reviewer comment: #1385 (comment).

Change `our_va` / `their_va` from "start of the aliased region" to "any address in the aliased region". Treating the fields as "must be in the region" rather than "must be the start" eliminates a failure mode in space_aware_map where a caller passes an unaligned address and the call silently produces a wrong link. Combined with the per-arch enum it lets every value of SpaceReferenceMapping have a well-defined semantic interpretation. Reviewer comment: #1385 (comment).

Make `space_aware_map` take `SpaceAwareMapping` and dispatch internally. Today the caller in Snapshot::new pattern-matches on SpaceAwareMapping and dispatches to either vmem::map (for ThisSpace) or vmem::space_aware_map (for AnotherSpace). Moving that dispatch inside the API — so callers hand the SpaceAwareMapping straight in and let the arch module route it — is more consistent with how walk_va_spaces already returns the enum, and it ties naturally into the per-arch typing change above. Reviewer comment: #1385 (comment).

Why bundle

Each change is small on its own but they share the same data shape (SpaceReferenceMapping and its fields). Landing them together means the walker/map API moves in one atomic step; doing them as four scattered patches would produce intermediate commits where the type lives in one place but the caller still dispatches externally, or the field semantics are partly updated. The whole cluster lives inside hyperlight_common::vmem + its arch modules and does not touch hyperlight_host at all.

Acceptance

• SpaceReferenceMapping defined in each arch module; on archs that never emit AnotherSpace, it is Void.
• depth replaced by an arch-specific enum.
• Doc invariant and implementations updated to accept any address in the aliased region.
• space_aware_map takes SpaceAwareMapping; the caller in Snapshot::new becomes a straight for sam in mappings { unsafe { vmem::space_aware_map(&pt_buf, sam, &built_roots) } }.
• i686 and amd64 both compile; existing snapshot unit tests still pass.

Links

• PR feat: i686 page tables, snapshot compaction, and CoW (standalone) #1385 threads this rolls up: r3132166982, r3132168766, r3132174870, r3132196047, r3132207257.