GPU implementation?

[thesamovar]

If huge amounts of RAM can be saved using #15, we could potentially fit the whole problem on a GPU and get large speed benefits. Further, since each cluster can be evaluated separately to each other cluster in the M and E steps, we could parallelise over multiple GPUs or computers in a cluster. The only information that would need to be communicated each iteration would be a few arrays of length the number of spikes per cluster, certainly less than 1 GB which could probably be transferred sufficiently quickly to make it advantageous.