Copy all eventspace counters to host efficiently at each time step

[denisalevi]

Currently, there are multiple kernels that require or could be optimized when having the number of spiking neurons on the host and each kernel calls their own cudaMemcpy or would have to.

Instead, implement an asynchronous cudaMemcpy from device to host after all thresholders (for any spike or event condition) are run.

1. Move the event counters for all eventspaces into a single contiguous array, such that one copy can copy all counters for multiple eventspaces at once. This would require changing the access to the counter variable in some kernels, but should be easy to do.
2. Detect from Python side which kernel will be the first to require the counter variables. That kernel should that synchronize with the asynchronous copy call. Also detect if no kernel requires the counters and don't copy in that case.

Related issues that would benefit or could be closed by this:

• Potential optimisations for paralelisation over spiking neurons for no_or_const_delay mode #146
• Remove duplicate cudaMemcpy every time step #149
• Call reset kernel only with as many threads as there are spiking neurons (not as there are neurons in total) #272
• Optimize PopulationRateMonitor #285