Potentially transfer to a tensor-based model rather than a DAG-based one

[HyperCodec]

The tensor-based model is much simpler and easier to understand and maintain, as well as eliminating performance costs in #99 and allowing GPU parallelization such as #60 to be implemented. Additionally, it would make #83 more performant and significantly easier to implement. However, doing so means the crate loses its uniqueness as one of the few DAG-based implementations. This will end up being another major rewrite if I decide to commit to it, though unlike the previous ones, it shouldn't have a major effect on the exposed API and requires little to no migration from users.

[HyperCodec]

This would also make mutation more performant, as I don't need all those cyclic checks

[HyperCodec]

After thinking about this a bit more, I don't believe this is a worthwhile change. GPU parallelization will still be a problem due to the freeform nature of the connections, which I am not willing to sacrifice in favor of a fully connected network or something of the sort. I'll leave this issue open just in case I come up with something else, but for now it's inevitably postponed.