Harden vGPU release ordering in the instance lifecycle

[yummybomb]

Summary

Layer 2 of the vendor VFIO vGPU stack (#366 ← this ← #363 ← #364 ← #321). Where #366 is a behavior-preserving refactor, this layer changes lifecycle semantics — release ordering and retention — still on mdev-era machinery, before the vendor VFIO backend lands above:

• Delete releases the vGPU before other teardown — a successful release persists the cleared assignment before later teardown. A failed release keeps the pre-refactor contract: it is logged and the delete continues, since the release guards never destroy a device they cannot prove is unowned and a skipped release is recovered by startup reconciliation. Stopped instances retry a previously failed release.
• Stop keeps its no-op contract — a failed retained-release on stop logs and retains the assignment metadata instead of failing the stop; assignments are passed to DestroyVGPU as a struct.
• Forks don't inherit retained assignments — a retained (failed-release) vGPU assignment is stripped from fork metadata so a fork can't release its parent's device.
• Start persists a stale release immediately — releasing a stale assignment before start writes the cleared metadata at once, so a later start failure can't leave metadata pointing at a device that's already gone.

Testing

• go build ./..., go vet clean
• go test -race targeted lib/instances lifecycle suites pass

---

Note

Medium Risk
Changes ordering and persistence of GPU teardown and restart-policy blocking during delete; incorrect behavior could leak vGPU slots or briefly restart instances mid-delete, though failures remain logged with reconciliation paths.

Overview
This PR tightens vGPU assignment handling across stop/start/delete, fork, and snapshot restore so metadata matches live host state and forks cannot touch a parent's device.

Delete now blocks restart policy before teardown (and reloads metadata so later saves do not undo it), and releases vGPU immediately after the VMM is dead—before network, passthrough devices, or volumes. A successful release is persisted before the rest of cleanup; failures stay best-effort and delete continues.

Stop no longer clears vGPU fields when DestroyVGPU fails; the assignment stays on disk for retry. Stop on an already-stopped instance and start both retry or flush stale releases (start persists the clear before allocating a new vGPU).

Forks (instance and snapshot) call clearStoredVGPUDevice so they keep GPUProfile but never inherit a retained path/UUID. Restore snapshot copies the instance's current vGPU fields from live metadata instead of resurrecting the snapshot payload.

Tests cover delete ordering, fork/snapshot clearing, restore stale vs retained assignments, and stop/start retry behavior.

^{Reviewed by Cursor Bugbot for commit 2ac1a4b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[yummybomb]

Trimmed over-engineering after an architectural review against the parent task (host kernel 6.8 / Ubuntu 24.04 support):

• 1c1f52e — dropped the HasMdev/IsAllocated() compatibility shim. VirtualFunction is internal and never serialized in an API response or persisted, so a single Allocated field is enough.

Kept as-is:

• the delete/stop behavior change (fail and retain assignment metadata when release fails) — orphan reconciliation only runs at startup, so failing the delete is the correct fail-closed behavior;
• the darwin DestroyVGPU unknown-framework error — it keeps the fail-closed rollback semantics identical on macOS, which the lib/instances tests exercise in CI.

Tests: go test -race ./lib/devices, ./lib/resources, ./lib/hypervisor/qemu, ./lib/hypervisor/cloudhypervisor, and the vGPU-targeted ./lib/instances tests all pass; the remaining lib/instances E2E failures require KVM/images and fail identically on an unmodified checkout.