On x86_64, this is always 128 bits, rendering it incompatible with AVX.
GCC reports the maximum enabled vector width (which is what we want) and Intel reports the maximum supported vector width (which is okay too).

This prevents possible deadlocks during reduction at the cost of an
additional communication.

the wrapper did not compile but since this was not a problem in the last two years it can be removed

constructing an empty std::set has almost no cost