feat(microvm): apply outbound NAT and host-memory admission on both VM runtimes

[Shine-neko]

Two host-side facilities existed as shared modules but were only wired into one of the two microVM runtimes. Both are now applied on both.

Outbound NAT. host_nat masquerades the guest supernet so a VM can reach external networks. It was called from Firecracker only, so a Cloud Hypervisor guest could serve inbound traffic but resolved nothing outbound. Cloud Hypervisor now makes the same call, in the branch that allocates the guest network — so it stays a no-op for a VM that declares no ports and therefore has no network at all. The rule set is idempotent and process-global, so this costs nothing after the first VM.

Host-memory admission. resources::check_host_memory refuses a deployment asking for more memory than the host has available. Cloud Hypervisor and containerd ran it; Firecracker had no memory gate, so an over-ask died on an opaque allocation failure at boot — and only after copying a full rootfs image per instance. The check now runs in start_vm before that copy. A refusal is terminal rather than a crash loop, since a retry cannot free memory.

Neither module needed changes: both were already runtime-agnostic, they were simply not called.

Docs: documented outbound access on the Cloud Hypervisor page (including the CAP_NET_ADMIN requirement and what happens without it), documented the memory gate on the Firecracker page, added both to the runtime comparison table, and corrected the Cloud Hypervisor page which described resources.requests.* as ignored — it is not used to size the VM, but requests.memory does drive admission control and takes precedence over limits.memory there.

Tests

764 unit tests pass; cargo fmt and cargo clippy --all-targets are clean. The underlying modules already have unit coverage — what changed here is where they are called from, which only an e2e run on a KVM-capable host with the runtime binaries would exercise. Those suites do not run in CI today.