Peer-to-peer inference engine for your claw: run one model across your own devices, each holding part of it on its GPU.
