Learning Multi-Agent Communication through Structured Attentive Reasoning