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; and
• LSGAN generator/discriminator composition facts.
• shared weighted-objective algebra connecting continuous-latent, discrete-latent, and adversarial objectives.

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.