Reduce number of threads in push kernel for heterogeneous delays in bundle mode

[denisalevi]

We are calling the push kernel with as many threads as the largest synapse group ((preID, postGroup) pair). But in bundle mode we only need as many threads as the maximum number of bundles in a synapse group. For the Brunel Hakim benchmark that is 41 threads vs. 1024 threads.

Maybe round up to next multiple of warp size.

And if implement parallel reallocation of spike queue cuda vectors, then more threads might be beneficial though.