Feature quantization

[Jegp]

Quantization works well with feed forward dynamics. Unfortunately, we used the functional API to implement recurrent layers, which makes it hard to optimize, since PyTorch doesn't recognize the layers.

1. One approach could be to implement custom quantization primitives, but that would take time. It could, however be necessary if we'd like to fully support quantization on the operations level.
2. Another idea would be to refactor the current recurrent implementations as an activation + 2 linear layers.

I like the second version, since that's exactly what a recurrent layer is. We could even go further and remove our recurrent functionals, since I suspect few people are using the functions directly and because it reduces maintainance. The current PR contains a suggestion on how the LIFRecurrentCell module could look like - and a test that demonstrates that it works.

I'm eager to hear your opinion @cpehle! May also be interesting for you, @ChFrenkel