Potential bug for CUDA reduction

[illuhad]

In the CUDA model's reduction, we omit __syncthreads() at the warp level:

CloverLeaf/src/cuda/context.h

Line 220 in 3e10ff9

if (offset > 16) __syncthreads(); // only need to sync if not working within a warp

I suspect that this is incorrect for post-Volta GPUs, where at least a __syncwarp() might still be needed. So I suspect that we would need something like

if(offset > warpSize/2)
  __syncthreads();
else
  __syncwarp();

This would also solve portability issues when running the code e.g. using AdaptiveCpp PCUDA on OpenCL devices or CPU devices, where the warp size might be smaller than 32.