Unify vGPU retention wrapping and dedupe test fixtures

[yummybomb]

Summary

Cleanup pass on the vendor VFIO vGPU lifecycle layer (#321), no behavior change:

• One retention mechanism in start. The start-path rollback for a failed vGPU create previously hand-built VGPUCleanupPendingError in two places instead of going through vgpuRetention like create does. The branch now records retention via markRetained and lets the existing deferred wrapPending produce the error, so the pending-error construction and the retained/unretained metric each live in one place. Error text, Retained flag, metric values, and wrapping order are unchanged.
• Inlined deferWrapPending. It was a one-line adapter around wrapPending; the ordering constraint it documented belongs at the two call sites, which now carry it.
• Deduped test fixtures. The identical 5-field vendor VFIO VGPUDevice literal appeared three times in vgpu_test.go; extracted testVendorVFIODevice.

Reviewed the rest of the diff for dead branches and redundant tests and deliberately left it alone: the per-callsite retention-stub rejection tests match the four independent guards, the reconcile tests each cover a distinct fail-closed path, and the interface{ StartVGPUReconciler } assertion in cmd/api/main.go follows the file's existing idiom for optional manager methods.

Testing

• go build ./..., go vet ./... clean
• go test -race passes for lib/devices, lib/hypervisor/qemu, and the lib/instances vGPU/retention/reconcile/lifecycle suites
• cmd/api/api and lib/instances integration failures in this environment are identical on the unmodified base branch (network/hypervisor privileges), per-test diff verified

[yummybomb]

Closing because this cleanup commit was pushed directly onto #321.