pvskand/3dpwm ? reverse-engineered prompt

Build me a Python project that implements the 3D Point World Models paper. I want it to train on partial point cloud data, complete the scene first, then learn action conditioned dynamics in 3D so it can do long horizon rollouts and score how good a plan is.

Please include a simple training flow and a test script that can load a checkpoint and run inference from a config file. It should work with the StackCube dataset setup described in the README, using the provided data layout and a single config file for the main experiment. Make it feel ready to run on Linux with Python 3.11 and an NVIDIA GPU, and include clear instructions for installing dependencies, downloading the dataset, training, resuming, and testing.

If you need to, look up the current docs for any tricky library details before wiring it up.