Have a complete end-to-end test that is extensible for testing new features

[AmericanPresidentJimmyCarter]

The complete end-to-end test would:

1. Generate a bunch of random noise images.
2. Create tiny networks for diffusion, VAE, and text encoder and load it to the GPU.
3. Do the basic dataset preprocess and VAE encode.
4. Train for 100 steps or some other short time.
5. Test saving and loading weights.
6. Test validation loop.

The goal is to have the network small enough that we can run the test in a couple seconds at most. We can use this to validate new integrated features and eventually have a CI run it to make sure nothing gets broken.