Multi-Architecture Expansion

• New model architectures: GeoLDM (VAE + latent diffusion), DiffLinker, GFMDiff, ADiT, TABASCO, Shepherd, and coordinate-only flow matching (flowmol)
• New regression backbones: eSEN and EquiformerV2
• Data pipeline: lazy-chunked/LMDB datasets, custom node featurization, graph compaction
• Training engine rewritten on PyTorch Lightning, with hyperparameter sweep support
• Generation: skeleton-based outpainting, structural guidance, XTB post-processing, NaN/outlier detection
• Unified MolCraftDiff CLI with optional-dependency handling for data/analyze extras
• New Sphinx docs, tutorials, and logo