feat(proxmox): give guest disks a dedicated IO thread

[v0l]

Problem

Every VM is created with scsihw: virtio-scsi-pci and no iothread, so all guest block IO is serialised through QEMU's main event loop — the same thread that services device emulation and networking. That caps a guest at roughly one thread's worth of storage throughput regardless of how fast the host pool actually is.

Measured on a production host (R740xd, 12x U.2 NVMe):

Layer	Throughput
Raw NVMe device, 1M blocks QD1	1.7 GB/s
ZFS zvol, 4M blocks	3.0 GB/s
8 parallel zvol streams	8.2 GB/s aggregate
In-guest	~250 MiB/s

The drives are healthy (2–3% wear after 4.3 years, 100% spare, 23°C) and the pool is at ~0.4% of its IOPS capability. The bottleneck was entirely the guest disk configuration.

On the affected host: 0 of 70 VMs had iothread=1, and 68 of 71 used virtio-scsi-pci, which cannot support it.

Change

• make_config now emits scsihw: virtio-scsi-single — the only SCSI controller type qemu-server accepts iothreads on.
• import_disk_image adds iothread=1 to the scsi0 device string at VM creation.
• apply_disk_limits → apply_disk_options, and it no longer returns early when a template has no throttle limits. Without that, VMs without limits would never converge onto iothread=1.

Rollout behaviour

Existing VMs pick up the new settings the next time their config is reconciled. A SCSI controller change only takes effect at VM start, so an already-running VM keeps its current behaviour until it is stopped and started — a guest-initiated reboot is not enough. No mass restart is triggered by this PR.

Mid-transition VMs are safe: qemu-server warns and ignores iothread on controllers other than virtio-scsi-single rather than erroring —

} elsif ($drive->{iothread}) {
    warn "iothread is only valid with virtio disk or virtio-scsi-single controller, ignoring\n"

So a VM whose config has iothread=1 but is still running the old controller simply ignores the flag until its next start.

Tests

• test_make_config_uses_virtio_scsi_single — guards the controller/iothread pairing from drifting apart.
• test_apply_disk_options_sets_iothread_without_limits — covers the removed early-return; asserts a PATCH is still issued and the bare volume ref is preserved.
• test_apply_disk_options_preserves_ssd_params — existing regression test, extended to assert iothread=1.

cargo test -p lnvps_api_common --features proxmox → 359 passed, 0 failed. cargo clippy and cargo fmt --check clean.

API impact

None — no request or response shape changed. Documented in API_CHANGELOG.md under Unreleased.