feat: Prototype virtio-mem auto memory

[lox]

spore run --memory auto currently pays the full 16GiB guest startup cost before PID 1 can respond, even for tiny commands. The goal of this prototype is to make auto boot like small memory while still exposing the 16GiB ceiling as elastic guest RAM.

Design: fresh managed runs are the only path that opts into virtio-mem. They boot with 512MiB of normal RAM, advertise the remaining 15.5GiB as one grow-only virtio-mem region, then request the plug after the guest readiness path. Explicit memory sizes, custom --kernel, SPOREVM_KERNEL_IMAGE, custom initrd, capture, and resume stay fixed-RAM so existing artifacts do not silently depend on hotplug. The virtio-mem device is shared virtio-mmio model code; HVF and KVM only do the backend-specific memory mapping behind the existing VM boundary.

This moves the managed run kernel default to cleanroom-kernels v0.6.0, and both the Zig resolver and shell helper verify the required hotplug config symbols before accepting a managed kernel.

Developer impact: the public --memory auto surface stays the same, but fresh managed runs can take the small-memory cold path while still growing to 16GiB after boot. This is intentionally a prototype: no pressure policy and no unplug policy.

Local checks already run: mise run test, mise run build, formatting/shell/diff checks, and HVF smokes. The managed /bin/true smoke reported ram_mib=512, virtio_mem_mib=15872, response_ms=31; the custom-kernel guard reported ram_mib=16384, virtio_mem_mib=0; the delayed Alpine smoke reported MemTotal: 16760636 kB after plug.

Kept draft until A1/KVM validation runs on the branch.