Host memory usage a bit excessive

[neworderofjamie]

I think we're a bit over-zealous in deep-copying stuff during model compilation and this leads to e.g. large EGP values being duplicated. I suspect deep-copying is the correct operation here but, retrospectively, allocating massive np.empty arrays in the model class is a bit silly - probably can just specify a size and allocate in the compiled model.

Only really found this trying to do DVS gesture inference on my laptop's 8GB RAM but probably a good thing to fix for future, larger datasets.