Pytorch firing parity

[williamnourse]

Discovered sources of firing rate discrepancy in PyTorch backend when compared to Brian2. Discrepancies were as follows:

1. Poisson inputs were being treated as incoming spikes, and filtered through the synaptic time constant. Fixed so that poisson inputs are directly injected into the membrane.
2. Poisson inputs were incorrectly triggering a refractory period in their stimulated neurons. Refactored the refractory period mechanism to be neuron-specific, allowing exclusion of stimulated neurons.

New firing rate statistics compared to ground truth are as follows, based on t_run = 10.0 and n_run = 1:
Ground truth: Brian2 (CPU)

  Brian2 (CPU): 168,903 spikes, 437 active neurons

  PyTorch (CUDA) (Original Version):
    Spikes:          95,052  (ratio: 0.5628)
    Active:             380   (Jaccard: 0.8442, shared: 374, gt_only: 63, backend_only: 6)
    Rate corr:   0.961903
    Median diff: 47.45%
    Verdict:     CLOSE

  PyTorch (CUDA) (Poisson inputs ->  membrane):
    Spikes:         174,681  (ratio: 1.0342)
    Active:             437   (Jaccard: 0.9729, shared: 431, gt_only: 6, backend_only: 6)
    Rate corr:   0.999439
    Median diff: 5.06%
    Verdict:     MATCH

  PyTorch (CUDA) (Poisson inputs -> membrane, Poisson inputs X refrafctory period):
    Spikes:         172,730  (ratio: 1.0227)
    Active:             435   (Jaccard: 0.9684, shared: 429, gt_only: 8, backend_only: 6)
    Rate corr:   0.99947
    Median diff: 3.62%
    Verdict:     MATCH