Probably don't need an intermediary type for serialization

[HyperCodec]

The NNTSerde struct makes it easier to write because of the derive macro, but it is also a lot slower than just converting directly from the neural network topology.

[HyperCodec]

The best solution for this would be to fix the Arc<RwLock> spam, although this may introduce a ton of complexity in the diversity between with and without the rayon feature

[HyperCodec]

If #60 does end up rewriting everything without Arc spam, this will become significantly easier.