Release minimal codebase for MSMD based on FLAME topology

[EvanSamaa]

---

Summary

This PR introduces a clean public release of the MSMD codebase, restructured from the original research code. The original was adapted from [DiffPoseTalk](https://github.com/DiffPoseTalk/DiffPoseTalk) and the internal ubisoft-laforge-msmd main branch. This release uses the FLAME topology with non-proprietary data, removes the dependency on the proprietary SEREP encoder, and makes the core training and inference pipelines runnable out of the box.

---

What's changed

• Replaced SEREP encoder with FLAME: all SEREP calls have been replaced with FLAME equivalents
• Added pre-trained model weights: enables direct comparison with MSMD in future work
• Reorganized into a standard ML package layout: msmd/models/, msmd/data/, msmd/utils/, msmd/options/, with data/ and pretrained_models/ as top-level directories
• Replaced all hardcoded absolute paths with relative paths; FLAME model files now resolve from pretrained_models/flame/
• Removed all sys.path.append hacks: the package is now properly installable and importable
• Fixed broken import in benchmark_Inference_DPT_runtime.py: was importing from refactor_train_dpt (nonexistent); corrected to msmd.data.dataset_factory
• Dropped unused expressive-speech2face imports from inference files
• Added standalone inference.py: runs on a single audio file with no dataset required; outputs motion coefficients as .npy or .json
• Cleaned up training scripts (train_diffusion.py, train_se.py): removed unused visualization imports, updated experiment output dirs to experiments/
• Added scripts/: train_msmd.sh, train_dpt.sh, train_se.sh updated with relative paths and plain bash (SLURM directives retained but commented out)
• Added README.md: covers installation, FLAME setup, dataset structure, training, and inference
• Added requirements.txt

---

Files not included in this release

• expressive-speech2face/ — spiral autoencoder, used only for visualization
• jam_mesh/, annotation_scripts/ — preprocessing only
• Visualization and user study scripts
• MSMD_venv/

---

Testing

• python -c "from msmd.models.diff_talking_head import get_difftalkinghead_model" passes
• python -c "from msmd.data.dataset_factory import create_datasets" passes
• python inference.py --help prints usage
• python train_diffusion.py --help prints usage