End-to-end trajectory planning with generative models

Diffusion policies and language-model planners both produce trajectories, and they fail in different ways. Implement both in our closed-loop CARLA setting and find out where each breaks.