fix(lcow/v2): report physically backed VM stats

[shreyanshjain7174]

Summary

• derive LCOW V2 controller memory backing from the HCS Memory.AllowOvercommit setting on cold boot and migration import
• avoid looking up a vmmem process for physically backed UVMs
• add focused coverage for the memory-backing decision

Problem

LCOW V2 physically backed sandbox stats had two failure modes:

1. the controller sourced isPhysicallyBacked from the stronger FullyPhysicallyBacked device annotation instead of !Memory.AllowOvercommit, so a non-overcommitted VM could report a zero working set;
2. Stats() unconditionally called LookupVMMEM before branching, even though physically backed UVMs use HCS AssignedMemory. When that lookup failed, containerd dropped the metric and callers observed an empty metrics response.

The change mirrors the established LCOW V1 behavior: memory is physically backed when overcommit is disabled, and vmmem working-set lookup is only used for VA-backed UVMs.

Validation

• go test -tags "windows lcow" ./internal/controller/vm/... -count=1
• go test -tags "windows lcow" ./internal/builder/vm/lcow/... -count=1
• rebuilt containerd-shim-lcow-v2.exe from this branch and ran Test_SandboxStats_WorkingSet_PhysicallyBacked/LCOW against a local build 26200 host: PASS
• restored the original deployed shim after the functional run

[rawahars]

Do we not need sandboxOptions.FullyPhysicallyBacked option?

[shreyanshjain7174]

Yes, we still need and retain sandboxOptions.FullyPhysicallyBacked. This change only stops using it as the stats predicate. Effective memory backing is !Memory.AllowOvercommit, while FullyPhysicallyBacked additionally controls device backing (for example, disabling VPMEM in favor of SCSI) and is still round-tripped in the migration payload. A caller can set AllowOvercommit=false without setting FullyPhysicallyBacked; that was the failing case. This mirrors V1's separate physicallyBacked and devicesPhysicallyBacked fields.

[shreyanshjain7174]

The key is whether the option remains required for device/VPMEM behavior even though it is no longer the right stats predicate.