Two major updates to the pipeline:
- Modify the data set generation to repeatedly hadronize the same parton-level event, such that we can train the diffusion model to learn the (forward or backward) one-to-many mapping between partons and hadrons.
- Include the Shift-DDPM model implementation to support conditional generation (e.g. generate hadron sample given a provided parton sample + noise) rather than traditional generation (e.g. generate hadron sample from noise).