Apply the persistant training schema form NCAs

[Francesco215]

This is similar to what is done in the neural CA paper

To do:

• [x] Define a pool of seed states to start the iterations from, initially filled with the single black pixel seed state.
• [x] Sample a batch from the pool for use in the training step.
• [x] Replace one sample in this batch with the original, single-pixel seed state to prevent the equivalent of "catastrophic forgetting".
• [x] Conclude the training step.
• [x] Replace samples in the pool that were sampled for the batch with the output states from the training step over this batch.