Implement RFDiffusion and RFDiffusion All-Atom in DeepChem

RFDiffusion (Watson et al., Nature 2023) is the state-of-the-art protein backbone diffusion model but DeepChem currently has no protein structure generative model. This project implements RFDiffusion and RFDiffusion All-Atom as native DeepChem TorchModel subclasses, covering a SE(3)-equivariant denoiser with Invariant Point Attention, a RoseTTAFold-style multi-track architecture, SO(3) diffusion for rotations, conditioning mechanisms (motif scaffolding, symmetry, binder design), the all-atom extension for side chains and small molecules, and a full evaluation pipeline. Deliverables include merged PRs for the featurizer, CATH dataset loader, core layers, and the complete model, along with benchmark results and a tutorial notebook.