Fix and improve summed variables implementation (including adding synapses to synapses support and start fixing spikegenerator)

[denisalevi]

PR for #49 [WIP]

We have three versions in the commits below that should be benchmarked against each other (see #49 discussion):

1. Original algorithm with synapses * targets threads in total (d2ff3ff )
2. One thread per synapse, each thread using atomicAdd on the target of its synapse in global memory to calculate summed variables (c54e64c)
3. Same as 2. but using cudaMemset to reset summed variables instead of thrust::fill (653d487)

EDIT
This PR turned out to touch a couple of issue at the same time:

• Summed variables are working now, using global atomicAdd on postsynaptic neurons. I opened Summed variables: Calculate partial sums in parallel in shared memory instead of using global atomicAdd #197 for potential further optimizations and Add a summed variable example to our benchmarks #198 for adding a summed variable example to our benchmarks.
• To have all summed variables test passing, I fixed the synapses to synapses issues (Add support for Synapses to Synapses #133).
• I needed the spikegenerator to work for some tests, so I started fixing Spikegenerator period argument not working #48, but some tests are still failing and some more tests are needed, this will happen in PR Fix spikegenerator.cu template #194.
• And this PR also refactored some additional code: Jinja template block names changed and get_array_name implementation now takes into account the different prefixes we use for device and host variables.

I did not do the benchmarking that I suggested above. This could be done in #197 if it turns out to be relevant.