Add a delay parameter to synapses

[laurentperrinet]

Transmission of a spike from one neuron to another often takes a given delay. It is most often accounted by the transduction delay along the axon and thus depends on the distance between neurons and of the axonal types (e.g width, myelin, ...).

To develop generic SNN models, adding the capability to have a delay between neurons would allow to develop more realistic models for which one would see the propagation in the hierarchy, or within a single layer. As such it is related to #94 - yet adding a delay would be a specific feature.

An alternative would be to consider changing the dynamical equation for the delay that would be approaching asymptotically that of a pure delay.

[cpehle]

Delay is indeed an important thing to consider. One way to solve this is by introducing a delay buffer between layers, that even works for recurrence with some work, but it is problematic from a memory efficiency perspective. This is what is done for example https://github.com/IGITUGraz/LSNN-official/blob/a9158a3540da92ae51c46a3b7abd4eae75a2bb86/lsnn/spiking_models.py#L282.

[Jegp]

We could of course add this to individual neuron dynamics, but what aobut adding a "delay" layer that simply caches spikes for a certain time period, depending on a matrix of delay times? Could look something like this:

model = SequentialState(
    LIFCell(),
    SynapseDelay(),
    ...
)

Where SynapseDelay would initialize with something like positive uniform values. This could also be used in recurrent layers, because we could just add it to the recurrent linear weighting,