Investigate & benchmark cold TCG boot-time optimizations (virtio-rng, random.trust_cpu, snapshot/restore)

First boot of the binder: vm TCG path took ~8 minutes on a no-KVM host (accel: auto correctly fell back to TCG). The QEMU launcher (vm/qemu.py ~252–276) uses -accel tcg,thread=multi,tb-size=1024 -cpu max and appends console=ttyS0 root=/dev/vda rw init=/init panic=1 mitigations=off — notably no RNG device and no random.trust_cpu=on, so Android init may stall on entropy (amplified under TCG).

This is an investigation, not a confirmed fix: benchmark cold-boot wall time with vs. without -device virtio-rng-pci and with random.trust_cpu=on, on the same host, several runs each. Separately, document/validate snapshot/restore (cli.py:snapshot/restore) as the warm-start path — boot once, snapshot the booted state, restore for near-instant subsequent starts.

Do not add vCPUs: the design log (vm-rnd-log §B.5) and resolve_smp already found -smp 4 optimal and -smp 8 regressing under MTTCG.

Acceptance: a short report with numbers on the RNG options, and a documented snapshot/restore warm-start workflow.

Start with vm/qemu.py around lines 252–276 and inspect cli.py's snapshot and restore commands. Benchmark cold TCG boots on the same no-KVM host with and without virtio-rng-pci and random.trust_cpu=on across several runs, without changing vCPU counts. Document the measured RNG results and a validated boot-once, snapshot, and restore workflow.