Avoid metadata writes when deleting instances without vGPUs

[yummybomb]

Why

#485 made DELETE return an error when vGPU release fails, instead of logging a warning and continuing. That was the right call for instances that actually hold a vGPU, but it exposed a pre-existing problem: delete called releaseStoredVGPUPersisted for every instance, GPU or not, and that helper always rewrites the metadata file via a temp file and rename.

For an instance with no vGPU that write does nothing useful. Before #485 its failure was swallowed, so nobody noticed. After #485 an ENOSPC on that write aborts the delete before any instance data is removed. On a full disk that means the only way to free space is blocked by the lack of space.

What changed

• Delete only calls releaseStoredVGPUPersisted when the instance has a vGPU assignment. This matches how start and the vGPU reconciler already gate the same helper.
• Real vGPU cleanup failures behave exactly as after Wait for exiting VMM tasks and retain failed vGPU cleanup #485: the error is returned, metadata and the GPU claim are retained, and network/device teardown does not run.
• Restart-policy handling is unchanged. The manual-stop marker is still written before teardown, and only when the instance has a restart policy.

Testing

• New TestDeleteWithoutVGPUDoesNotRewriteMetadata asserts the metadata file keeps the same inode through delete for a non-GPU instance. It fails on main and passes here.
• Targeted delete, vGPU release, lifecycle, and reconciliation tests pass with -race; the new test and the existing vGPU delete tests were repeated 25 times.
• Ran the real delete path against stopped Firecracker metadata with no GPU on an isolated 8 MiB tmpfs under both inode and block exhaustion. Unpatched code fails and leaves the instance behind; patched code deletes it.

Not covered: no hypervisor was started and the full instances suite was not run. Instances that carry a restart policy still need one metadata write during delete to block auto-restart, so a full disk can still fail those; none of our callers set a restart policy today.

---

Note

Medium Risk
Touches instance delete teardown ordering for vGPU vs non-vGPU paths; wrong gating could skip needed GPU cleanup or regress disk-full delete behavior.

Overview
Delete no longer calls releaseStoredVGPUPersisted for instances without a vGPU assignment, aligning delete with how start already gates that helper. Previously every delete attempted a metadata rewrite (temp file + rename) even when there was nothing to release; after #485 a failed write (e.g. ENOSPC) could abort deletion before any instance data was removed.

When a vGPU is assigned, behavior is unchanged: release runs inside the same guard, failures still return an error and retain metadata for retry, and network/device teardown still does not run on failure.

Adds TestDeleteWithoutVGPUDoesNotRewriteMetadata, which checks the metadata file keeps the same inode through delete for a non-GPU instance (via a hook during device detach).

^{Reviewed by Cursor Bugbot for commit 50b4444. Bugbot is set up for automated code reviews on this repo. Configure here.}