Attempt cache thrashing fix for low synapse count vector access

[otcathatsya]

• This is an attempt to fix an observed niche issue where, on a C++ level, connection creation will be slowed down by the num_connections_ vector accessing inner indices on the same cache line and therefore invalidating memory across threads, usually when there is only a few synapse models (e.g. when running with -Dwith-modelset=iaf_minimal)
• Occurs mostly commonly with jemalloc:

Allocations are packed tightly together, which can be an issue for multi-threaded applications. If you need to assure that allocations do not suffer from cacheline sharing, round your allocation requests up to the nearest multiple of the cacheline size, or specify cacheline alignment when allocating.

• Very much experimental, this fixes the time_construction_connect difference in SLI tests but makes no difference on a Python level (neither did the problem occur when timing from Python to begin with. Too insignificant?)
• This does increase memory consumption up to a certain point (the L1 cache line usually is 64 bytes), but stays below full modelset levels; it is actually lower when running the full modelset with this alignment.
• Maybe the destructive interference constexprs can be useful for something else further down the line
• If this does get considered it should probably be amended to include -Wno-interference-size to suppress the compile warnings about this value varying between platforms, or be an optional CMake flag altogether
• To-do: find a way to apply the destructive interference offset to the outer vector for more memory savings

[github-actions]

Pull request automatically marked stale!

[terhorstd]

@suku248 and @heplesser want to be in the loop for this.