rlx-distributed 0.2.18

RLX multi-node pipeline-parallel execution: graph partition into weight-sharded stages + tensor transport, to pool RAM across machines for models larger than any single node.