Learning Multi-agent Behaviors from Distributed and Streaming Demonstrations