Tunable GPU block sizes

[JPRichings]

Noticed that

QuEST/quest/src/gpu/gpu_kernels.cuh

Line 50 in 9d7618d

const int NUM_THREADS_PER_BLOCK = 128;

is fixed for all target hardware and is a bit large for common tuning recommendations.

Plan to change this to allow a compile time default and a setter-getter interface to allow performance tuning tests.

[otbrown]

As discussed, will likely do something with an env var to be symmetric with OMP_NUM_THREADS, but should be simple enough!

(Famous last words.)

[otbrown]

Mere minutes later I realised that would be a bad idea, as we have an external source of truth (the OpenMP runtime) for OMP_NUM_THREADS, but not for NUM_THREADS_PER_BLOCK...