Latent generative model theory

This entrypoint collects the proved theory facts for TorchLean's latent generative model specs:

• VAE reparameterization and β-VAE objective decomposition;
• VQ-VAE codebook lookup and loss decomposition;
• LSGAN generator/discriminator composition and target-score facts.

The executable model equations and the heavier probabilistic/game assumptions stay separate. These files prove the stable rewrite and optimization facts that examples, verifiers, and theory modules can use without unfolding the model specs by hand.