Learning Distributional Demonstration Spaces for Task-Specific Cross-Pose Estimation