Neural Mixer: Evaluate, refactor or redesign where necessary, unit test, and document

[fromdavelewis]

[paxcema]

As reported by @tomhuds we need to take a closer look at the efficiency after selecting initial hyperparams, seems to scale poorly.

[tomhuds]

@paxcema can we close?

[paxcema]

Still relevant, will be worked on by @pafluxa and I at some point in the coming weeks.

[paxcema]

Ok, closing now. We can still improve but the mixer is at a better place in terms of speed/accuracy tradeoff now.