Integrate private copy-on-write mappings with guest memory backing

Tracking issue: MicroVM support and private snapshot restore.

Motivation

Multiple VMs should be able to start from the same immutable RAM image while keeping
their writes private. A host mapping primitive alone is insufficient: guest memory
backing must consume it correctly and expose the accounting needed to understand
faults and private-memory behavior.

Proposed Scope

• Add private copy-on-write file views to sparse_mmap on the supported hosts.
• Integrate those views into membacking, including ownership, lifetime, and fault
  accounting.
• Validate mapping offsets, lengths, alignment, and address arithmetic before
  creating or exposing guest memory.
• Keep the host primitive and its working memory-backing consumer in one PR.
• Define errors and capability limits without silently changing private mappings
  into writable shared mappings.

Acceptance Criteria

• Two private views initially observe the file contents, but writes to one
  view do not modify the file or the other view.
• A guest-memory-backing test exercises the private mapping through its actual
  consumer, not only the low-level mapping API.
• Invalid ranges and mapping failures return errors without leaked mappings or
  partially exposed guest memory.
• Fault accounting follows the documented meaning and does not mix private and
  shared mapping modes.
• Supported Windows and Linux paths have focused coverage, with unsupported
  combinations reported explicitly.

Dependencies

No feature prerequisites. Snapshot publication and restore will consume this
memory capability in a separate PR.

Non-Goals

This issue does not define a snapshot manifest, capture transaction, restore-memory
expansion policy, or lazy hypervisor registration.

Start by reading the sparse_mmap and membacking components, then trace the existing host mapping consumer and supported Windows and Linux paths. Use the listed acceptance criteria to define completion: private views must isolate writes, validate ranges, clean up failures, and report fault accounting and unsupported combinations explicitly.