Diffusion theory

This entrypoint collects the diffusion-theory facts that connect TorchLean's executable sampler specifications to the mathematical language used in diffusion and score-based generative modeling.

This entrypoint collects:

• ForwardGaussian: a mathlib-backed result showing that affine forward noising of a standard Gaussian remains Gaussian.
• Samplers: proved boundary, dynamics-adapter, and Euler-stability facts for DDPM, DDIM, and probability-flow samplers.
• ImageDDIM: equalities connecting the public image helpers to their coefficient indexing, denominator floor, and clipped reconstruction specifications.

Probabilistic claims and executable sampler claims stay separate. The spec layer defines the noising and reverse-update functions; this theory layer records the mathematical facts we can prove cleanly about those definitions.