Fix spikemonitor for subgroups

[denisalevi]

Fixes #286

This implementation is a quick solution that does the following when recording with SpikeMonitor from a Subgroup.

• In Brian2CUDA, each NeuronGroup has its own spikespace, which is filled with the neuron IDs of spiking neurons. In cpp_standalone, this spiekspace is sorted by neuron IDs. In Brian2CUDA, it is not, which is why we can't just read a consecutive subsection of the spikespace, where all neuron IDs belong to one Subgroup.
• Instead, this implementation calls thruds::count_if to get the number of spiking neurons in the Subgroup and then calls thrust::copy_if to copy those IDs into another array (spikespace for subgroup).
• This means we can use the same spikemonitor kernel for NeuronGroup and Subgroup recordings, but passing a different spikepace pointer.

This implementation is surely not the most performant, but was instead easy to implement. I added #293 with suggestions for a better implementation.