directly save encoder and decoder models, as well as weights

[quantumjot]

Here: https://github.com/lowe-lab-ucl/cellx-predict/blob/9e43599973c77b436cfb637ab1dcdf68f1ce2a56/cellxpredict/train.py#L57

we should also be able to save a model (rather than just weights, since https://github.com/quantumjot/cellx/pull/44 fixed the custom layer serialization). This would make loading and inference much easier...