feat(datasets): add training data preprocessing (GameFactory example)

[jankinf]

• Add hy_preprocess/ with preprocessing scripts for training data.
• GameFactory pipeline: preprocess_gamefactory_dataset.py converts GameFactory/Minecraft data into the format expected by CameraJsonWMemDataset (pose JSON, action JSON, VAE latents, temporal alignment).
• Utilities: generate_neg_prompt_pt.py for negative-prompt generation; shell scripts under scripts/ to run preprocessing and neg-prompt generation.
• Supports configurable resolution, frame count (e.g. 129 frames), and optional subsampling for quick runs.

This is provided as a reference implementation; adapt paths and options for your own datasets and environment.