Replaced global cuda_callable lambdas with functors to avoid lots of nvcc warnings

Functors can also be better optimized by the host compiler, because nvcc sucks...