Modelplane v0.7 provisions inference clusters on CoreWeave, and fixes the rough edges we found running real fleets ahead of KubeCon.